Abstract: With the wide attention in the field of cloud computing, the cluster container orchestration management platform Kubernetes has been widely used in service scenarios such as automatic deployment and release of container application service, elastic expansion of application and rollback update, and fault detection and self-repairing. Fifth-generation reduced instruction-set computer (RISC-V) includes four technical characteristics and advantages: fine simplification, modularization, scalability, and open source, and it has attracted extensive attention from academia and industry. Based on the collaborative research of Kubernetes ecology and RISC-V ecology, this study supports scheduling tasks with heterogeneous instruction set architecture (ISA) for the Kubernetes scheduler. Through the quantitative analysis of various computing task requirements of RISC-V instruction set architecture in the production environment, it is found that the existing Kubernetes cannot schedule the computing tasks of RISC-V instruction set architecture, and in particular, it fails to employ the extended instruction set architecture characteristics defined by RISC-V developers to provide high-performance and reliable services. In order to solve these problems, this study proposes an ISAMatch model which comprehensively considers the affinity of the instruction set, the number of nodes in the same instruction set architecture, and the utilization of node resources, so as to realize the optimal allocation of tasks. Based on the existing cluster scheduler, this study improves its scheduling requirements for multiple instruction set architecture tasks. In addition, compared with the default scheduler whose accuracy rate is 62% (scheduling RISC-V basic instruction set tasks), 41% (scheduling RISC-V extended instruction set tasks), and 67% (scheduling RISC-V instruction set tasks with “RISC-V” node matching label), ISAMatch model can achieve a task scheduling accuracy of 100% without considering resource constraints.
Abstract: The continuous evolution of hardware and software technology demands higher execution performance from instruction set architecture emulators represented by QEMU. This study analyzes the limitations of QEMU’s existing dynamic jump handling mechanism in the scenario where the emulated architecture supports virtual memory, and proposes an optimized scheme based on address space identifiers suitable for common virtual memory systems. The proposed scheme is implemented for the RISC-V frontend in QEMU mainline 6.2.0 version. Evaluation results show that the dynamic jump scheme based on the address space identifier achieves an average performance improvement of 12% compared to the native QEMU.
Abstract: As a typical example of reduced instruction sets, RISC-V can also reflect some disadvantages of the reduced instruction set computer (RISC), and large program size is one of the problems. Compared with the complex instruction set computer (CISC), RISC generally requires more instructions to implement complex operations and results in a large binary size of the program. Meanwhile, RAM and ROM in embedded devices are generally small. Therefore, it means that the binary size of the program is significant for embedded scenarios. In view of this, the Zce sub-extension of RISC-V has developed a series of instructions to reduce the program size as much as possible. Specifically, the instructions represented by the LWGP are used to reduce the number of instructions when loading/storing bytes. This study analyzes the principle of the LWGP instructions in reducing the code size and implements it on the LLD linker. It also evaluates the efficiency of LWGP in reducing the binary size of the program by analyzing the change in program size before and after using LWGP instructions and puts forward recommendations for improvement.
Abstract: In recent years, as RISC-V architecture spreads rapidly in the industry due to its advantages of open source, concision, and modularization, massive processor IP cores and system on chip (SoC) based on the RISC-V architecture have emerged in the market. The existing debuggers serve as an important tool in developing RISC-V software, but they face low performance, high deployment cost, and high difficulty in secondary development and struggle in meeting the needs of RTL design and verification, software development and debugging, and mass production/batch programming of RISC-V architecture-based chips. To solve these problems, this study proposes a new, open-source, and modularized RISC-V debugger protocol stack design scheme based on the lightweight remote procedure call—Morpheus. Experiments and analysis results have shown that this debugger protocol stack can effectively reduce the deployment cost and the difficulty of secondary development and improves the debugging performance.
Abstract: Memory safety is critical but vulnerable. In view of this, numerous defense countermeasures have been proposed, but few of them could be applied in a production-ready environment due to unbearable performance overhead. Recently, as open-sourced architectures like RISC-V emerge, the extension design of enhancing hardware memory safety has revived. The performance overhead of hardware-enhanced defense techniques becomes affordable. To support the extension design of enhancing memory safety systematically, this study proposes a comprehensive and portable test framework for measuring the memory safety of a processor. In addition, the study achieves an open-sourced initial test suite with 160 test cases covering spatial and temporal safety of the memory, access control, pointer and control flow integrity. Furthermore, the test suite has been applied in several platforms with x86-64 or RISC-V64 architecture processors.
Abstract: This study expounds on the practice of embedded technology by using the AHL-CH32V307 hardware system based on the CH32V307 microcontroller with a RISC-V processor core from Nanjing Qinheng Microelectronics. Firstly, the study briefly introduces the knowledge system of the embedded system, reduces the development work of the embedded system with a high-tech threshold, and realizes the agile development ecosystem of embedded artificial intelligence. Then, the embedded development hardware is given and tested. With the intuitive experience of compiling, downloading, and running the first embedded program in a multi-functional embedded integrated development environment, students can start their journey to learn embedded systems. In terms of the hardware system corresponding to the development kit, this study describes the basic principle, circuit connection, and programming practice of some common controlled units in the embedded system, such as color lights, infrared sensors, and the tree structure of assembly project. Furthermore, a simple and practical embedded object recognition system based on image recognition is designed by using the CH32V307 microcontroller, and it can be used as a quick-start system of artificial intelligence. The teaching cases in this study are applicable to the teaching or technical training of embedded systems in colleges and universities, and they can provide a reference for application technicians in developing embedded systems.
Abstract: In the process of automated crane operations for port containers, the detection of container truck heads is an indispensable link. To solve the problem of low efficiency by manual confirmation and high costs and complex systems by the laser scanning method, this study proposes an algorithm based on video images of operation scenes and deep learning for target detection of container truck heads. Specifically, upon the construction of a sample data set of container truck heads, the DCTH-YOLOv3 detection model is used, and sample training is performed through the method of model migration learning. The DCTH-YOLOv3 model is an improved YOLOv3 model proposed in this study. The algorithm improves the FPN structure of YOLOv3 and proposes a new feature pyramid structure—AF_FPN. During the fusion of higher- and lower-order features, the AFF module with the attention mechanism is introduced to focus on effective features and suppress interference noise, which increases the accuracy of detection. In addition, the metric CIoU loss is used to replace L2 loss to provide more accurate boundary box change information and further improve the model detection accuracy. The experimental results indicate that the detection rate of DCTH-YOLOv3 can reach 46 fps on GTX1080TI, which is only 3 fps lower than that of YOLOv3. The detection accuracy can reach AP0.5 0.9974 and AP0.9 0.4897, in which AP0.9 is 16.4% higher than that of YOLOv3. Compared with the YOLOv3 algorithm, the proposed algorithm has higher accuracy and can better meet the requirements of automatic operations for high accuracy and fast identification in the anti-collision detection of container trucks.
Abstract: Although direct-current (DC) charging piles are effective power supply equipment for electric vehicles (EVs), their frequent faults pose a threat to the charging safety of EVs. Accurately predicting charging pile faults can effectively ensure the safety of EVs in the charging process. For this reason, a fault prediction model for DC charging piles based on an improved gated recurrent unit (GRU) is proposed in this study. Specifically, the common fault types of DC charging piles during charging are analyzed. Considering the small sample size of specific fault data in the actual collection, variational autoencoder (VAE)-based data augmentation is performed to expand the sample data. Then, on the basis of the current fault prediction method based on the GRU network model, this study resorts to the particle swarm optimization (PSO) algorithm to optimize GRU network parameters, employs the support vector machine (SVM) model to improve the classification function output by the network, and thereby proposes a PSO-GRU-SVM fault diagnosis model for DC charging piles. Finally, an example is discussed to compare the prediction accuracy before and after the improvement, and the confusion matrix heatmaps are comparatively analyzed. Furthermore, the proposed model is compared with two commonly used network models. The results show that the proposed method can effectively improve prediction accuracy and thus verify the feasibility of the proposed method.
Abstract: To effectively improve the transmission rate and reduce bandwidth burden, this study proposes a multiple-image encryption scheme based on compressed sensing and a hyperchaotic system. Specifically, several original images are spliced into a new plaintext image, and some plaintext information is combined with random positive integers to generate the initial value of the chaotic system. The pseudo-random sequence generated by the hyperchaotic system is utilized to produce the measurement matrix, scrambled sequence, and diffusion sequence the encryption process needs. Then, discrete wavelet transform, thresholding, and parallel measurement are performed to compress the plaintext image, which can effectively reduce the amount of operation data and greatly speed up the operation speed. Finally, the final ciphertext image is obtained by non-repeated scrambling and bidirectional mode-adding diffusion. Multiple levels of simulation experiments verify that the proposed algorithm can effectively resist cropping attacks and offer high security.
Abstract: With the increasing number of electric vehicles (EVs), the related supporting facilities are also facing great challenges. Unreasonable charging resource allocation will cause overcrowding at some charging stations during the peak charging period and affect the stable operation of power grids. A scheduling model considering multi-objective optimization is proposed. Upon the analysis of the queuing time of different charging options at the charging stations, a dynamic pricing model considering the queuing rate and time-of-use tariff is presented to affect the charging behavior of EV owners. The charging cost is calculated with the dynamic pricing model and the charging demand. Considering the travel time of the total charging path based on the starting and ending points, the optimization objective is to minimize the total cost, which is solved by the DEB-ABC algorithm. The simulations of 1 500 EVs in a certain area indicate that the proposed optimal scheduling model can reduce the waiting time for charging, charging costs, and total driving time and improve the utilization of charging stations in the area.
Abstract: As an innovative distributed ledger technology, a Blockchain has broad application prospects in many industries due to its features of decentralization, traceability, and tamper resistance. However, the existing single-chain structure of Blockchains faces problems such as low concurrency and high latency. The emergence of a new ledger technology based on the directed acyclic graph (DAG) structure is expected to break through the performance bottleneck of traditional Blockchains, but the current consensus mechanism based on the DAG-based Blockchain system is not mature. This study improves the security problems in the open representative voting (ORV), a consensus mechanism of the Nano network for the typical DAG-based Blockchain system, and proposes a consensus mechanism of open election representative voting (OERV) based on the representative election model. The rights and interests of the main representative nodes are dispersed; the degree of decentralization is enhanced, and the network security is improved. The experimental results reveal that the OERV algorithm has high performance and can enhance the stability and security of the system without sacrificing system efficiency. It is of practical significance for promoting the research on the consensus mechanism of DAG-based Blockchains.
Abstract: The varieties of tomato leaf diseases are of small differences and are hard to distinguish with the naked eye. Given that the classical convolutional neural network is exposed to various problems, such as a large number of parameters, heavy computation burden, a low identification rate of the model, and a large prediction error, this study proposes a disease identification method based on an improved MobileNetV2 network. A channel and a spatial attention mechanism are added to the right network layer to enhance the ability of the network to specify the features of diseased leaves and reduce the interference of irrelevant features. The Ghost module is used to replace some of the inverted residual blocks in the original model to ensure the accuracy of the model and reduce the number of parameters. The LeakyReLU activation function is employed to retain more positive and negative feature information in the feature map and thereby enhance the robustness of the model. Ten tomato leaf diseases, including early blight, late blight, spot blight, bacterial canker, erythema tetranycariasis, leaf mildew, and bacterial spot, are selected from the public dataset PlantVillage to serve as the experimental dataset. The experimental results show that the classification accuracy of the improved MobileNetV2 network reaches 98.57%, which is 2.29% higher than that of the original MobileNetV2, and the model size is reduced by 22.52%, representing a remarkable optimization effect.
Abstract: To address the problems of low contrast between target and background and lack of target feature information facing satellite video images, this study proposes a target segmentation and tracking method combining target motion information, spatio-temporal background, and appearance model. After the target area is obtained by positioning in the first frame, the histogram of oriented gradient method is employed to extract the features of the target, and the kernel correlation filter (KCF) is utilized to obtain the target tracking area 1. Subsequently, color and spatial features are used to build a spatial model of the context information about the target and its surrounding area and thereby obtain the target tracking area 2. Then, the visual background extraction algorithm is applied to detect the moving target in the target area in pixels and further obtain the segmentation area 3 of the single target. Finally, the correlation of the three areas is calculated, respectively, to obtain the optimal area as the final target tracking position and the template update sample. The experimental results show that compared with the KCF algorithm, the proposed algorithm obtains a significantly higher tracking success rate and accuracy and also achieves single target segmentation.
Abstract: Object tracking, a basic problem in computer vision, has a wide range of application scenarios. Due to the advance in the computational capacity of hardware and deep learning methods, conventional deep learning methods for object tracking have higher precision, but they face the problems of massive model parameters and high demand for computational resources and power consumption. In recent years, with the booming development of unmanned aerial vehicle (UAV) and Internet of Things (IoT) applications, a great deal of research focuses on how to achieve real-time tracking in embedded hardware environment with limited storage space and computational capacities and low power consumption. Firstly, object tracking algorithms in the embedded environment, including the ones combining correlation filters with deep learning and those based on lightweight neural networks, are analyzed and discussed. Secondly, deployment procedures of deep learning models and classical embedded object tracking applications, such as those in UAVs, are summarized. Finally, future research directions are given.
Abstract: Yak grade evaluation is an important part of high-efficiency yak breeding. To reduce the influence of imbalanced data set distribution on the prediction results of yak grading in the research, this study proposes a yak grade evaluation model based on an improved conditional generative adversarial network model, called VAE-CGAN. Firstly, to obtain high-quality generated samples, the model reduces the uncertainty from random variables by introducing a variational autoencoder to replace the random noise in the input of the conditional generative adversarial network. In addition, the model inputs the yak label as conditional information into the generative adversarial model to obtain the generated samples of the specified category, and the generated samples and training samples are utilized to train the deep neural network classifier. The experimental results show that the overall prediction accuracy of the model has reached 97.9%. The Precision, Recall, and F1 value on the grade prediction of premium yak have increased by 16.7%, 16.6%, and 19.4% respectively compared with those of the generative adversarial network. The results indicate the model can achieve yak classification with high accuracy and low misclassification rate.
Abstract: Present image encryption algorithms ignore the protection of the visual security of encrypted images. Therefore, it is valuable to combine a new cosine chaotic map (CCM) with Bayesian compressive sensing (BCS) and thus propose a visually meaningful image encryption (VMIE) algorithm. Firstly, a new one-dimensional chaotic map based on the cosine function is proposed to construct a controlled measurement matrix. In addition, the proposed new CCM can better disrupt the strong correlation of images. Secondly, the wavelet packet coefficient matrix of a plain image is scrambled by 2D Arnold scrambling algorithm. Then, the scrambled secret image is compressed and encrypted by a chaotic measurement matrix and bidirectional modulo-adding diffusion strategy. Finally, a visually meaningful ciphertext image is obtained by embedding the secret image into the carrier image after game-of-life (GOL) mixed scrambling through the least significant bit embedding algorithm. Simulation results and security analysis show that the proposed algorithm is feasible and efficient on the premise of ensuring visual security and decryption quality.
Abstract: With the development of the Internet, how to quickly obtain core information from massive news and make browsing easy has become an urgent problem for information departments. The existing TextRank and its improved algorithm fail to consider text features comprehensively in extracting news summaries. In selecting summaries, they only focus on the redundancy and ignore the diversity and readability of the summaries. In order to solve the above problems, this paper proposes a multi-feature automatic text summarization method, namely, MF-TextRank. A more comprehensive text feature information is summarized according to the structure, sentences, and words of news, which is used to improve the weight transfer matrix of the TextRank algorithm and make the sentence weight calculation more accurate. Furthermore, an MMR algorithm is used to update sentence weight, and the candidate summary set is obtained by beam search. According to the MMR score, the candidate summary set with the highest cohesion is selected as the final summary for output. The experimental results show that the MF-textRank algorithm outperforms the existing improved TextRank algorithm in extracting summaries and effectively improves the accuracy in this regard.
Abstract: In the application scenario of autonomous driving, YOLOv5 is applied to target detection, and the performance is significantly improved compared with that of previous versions. However, the detection accuracy is still low in the case of high running speed. This study proposes a vehicle-side target detection method based on improved YOLOv5. In order to address the issue of manually designing the initial anchor box size in training different datasets, an adaptive anchor box calculation is introduced. In addition, a squeeze and excitation (SE) module is added to the backbone network to screen the feature information for channels and improve the feature expression ability. In order to improve the accuracy of detecting objects of different sizes, the attention mechanism is integrated with the detection network, and the convolutional block attention module (CBAM) is integrated with the Neck part. As a result, the model can focus on important features when detecting objects of different sizes, and its ability in feature extraction is improved. The spatial pyramid pooling (SPP) module is used in the backbone network so that the model can input any image aspect ratio and size. In terms of the activation function, the Hardswish activation function is adopted for the entire network model after the convolution operation. In terms of the loss function, CIoU is used as the loss function of detection box regression to solve the problems of low positioning accuracy and slow regression of the target detection box during training. Experimental results show that the improved detection model is tested on the KITTI 2D dataset, and the precision of target detection, the recall rate, and the mean average precision (mAP) are increased by 2.5%, 5.1%, and 2.3%, respectively.
Abstract: Currently, patients are transported mainly by fuel vehicles. In view of this, this paper carries out a study to model patient transportation by electric vehicles and analyze the calculation examples of patient transportation by fuel and electric vehicles through comparison, so as to verify the feasibility and superiority of patient transportation by electric vehicles. Firstly, a mathematical model of patient transportation by fuel vehicles is constructed, which considers constraints such as the longest riding time of each patient, the maximum average speed of vehicles, and the time window for patients, with a goal of minimizing the sum of consumption and refueling costs of fuel vehicles. Secondly, a mathematical model of patient transportation by electric vehicles is constructed, which takes constraints such as the charging time of electric vehicles, the remaining power, the maximum average speed of electric vehicles, the longest riding time of each patient, and the time window for patients into account, with a goal of minimizing the sum of consumption and charging costs of electric vehicles. Finally, an example is selected and solved by LINGO software through programming to verify the feasibility and effectiveness of the mathematical models.
Abstract: Vehicle detection is an important research direction for intelligent transportation systems. In terms of vehicle detection from the monitoring perspective, a vehicle detection method based on an improved YOLOX algorithm is proposed. The YOLOX_S model with a smaller network depth is used to improve the network structure. The GHOST depthwise separable convolution module is adopted to replace some traditional convolutions, and model parameters are reduced with the model detection accuracy ensured. The CBAM attention module is integrated into a feature extraction network, and a feature enhancement structure is added to enhance the semantic information of feature maps obtained by the network and strengthen the ability of the network in detecting targets. By using the CIoU_loss to optimize the loss function, this paper finds that the positioning accuracy of the bounding box of the model is improved. The test results show that the detection accuracy of the improved network is increased by 2.01%, reaching 95.45%, which proves the feasibility of the improved method.
Abstract: Glioma is one of the most lethal tumors in the world. It is a malignant disease with high mortality, easy recurrence, and great harm to the body. At present, magnetic resonance imaging (MRI) technology, due to its characteristics of clear imaging effect and sharp contrast between different soft tissues, has become a commonly used medical method to diagnose patients with glioma. Given the lack of original glioma data set, this study, in cooperation with Liaoning Tumor Hospital, analyzed MRI images of 300 glioma patients in the hospital. The original glioma data set is established by classifying and further grading the original data through lesion determination, lesion location, and lesion qualitative. Analysis and experiment are conducted to verify its subsequent application. It is proved that the original data set can be used for image classification and segmentation, providing image data for tumor growth and reconstruction and sufficient help for clinical research and application of glioma.
Abstract: Original Harris hawks optimization (HHO) has low convergence accuracy and slow convergence speed and is easy to fall into local optimum. In view of these problems, an improved HHO based on a hybrid strategy (HSHHO) is proposed. Firstly, the Sobol sequence is introduced in the population initialization stage to generate a uniformly distributed population, which enriches the diversity of the population and helps to improve the convergence speed of the algorithm. Secondly, the limit threshold is introduced to make the algorithm perform global exploration when it does not obtain a better value within a certain number of iterations. This can improve the ability of the algorithm to jump out of a locally optimal solution and solve the problem that HHO is prone to fall into a locally optimal solution in late iterations because it only executes the development phase. Finally, a dynamic backward learning mechanism is proposed to improve the algorithm’s convergence accuracy and ability to jump out of the local optimum. The proposed algorithm is tested by nine benchmark functions and six CEC2017 functions and compared with various optimization algorithms and HHO variants. As a result, this study verifies the effectiveness of the proposed strategies and performs Wilcoxon signed rank test, Friedman test, and Quade test. The experimental results show that HSHHO has great performance in terms of convergence speed, optimization accuracy, and statistical tests. Furthermore, the proposed algorithm is applied to the design optimization of welded beams. The results show that HSHHO also has a positive effect on practical engineering optimization problems with constraints.
Abstract: The feature matching algorithm based on deep learning can produce larger scale and higher quality matching than the traditional algorithm based on feature points. This study aims to obtain a wide range of clear pavement crack images and solve the problem of missing matching pairs in weak texture image mosaics. The road image mosaic is realized based on the deep learning LoFTR (detector-free local feature matching with Transformers) algorithm. Given the characteristics of road images, the local mosaic method is proposed to shorten the running time of the algorithm. Firstly, the segmentation of adjacent images is conducted, and the dense feature matching is produced through the LoFTR algorithm. Secondly, the homography matrix value is calculated according to the matching results and the pixel conversion is realized. Thirdly, images after local mosaics are obtained through the image fusion algorithm based on wavelet transform. Finally, some images that are not input into the matching network are added to get the complete mosaic result of adjacent images. The experimental results show that, compared with methods based on SIFT (scale-invariant feature transform), SURF (speeded up robust features), and ORB (oriented FAST and army), the proposed method has a better effect on road image mosaic and higher confidence of matching results in feature matching stage. For the mosaic of two road images, the time consumed by the local splicing method is shortened by 27.53% compared with that before the improvement. The proposed mosaic scheme is efficient and accurate, which can provide overall disease information for road disease monitoring.
Abstract: For the severe challenges brought by the fluctuation and randomness of photovoltaic power generation to the load prediction of the dispatching department and the safe operation of the power grid, this study proposes a photovoltaic power prediction method of bidirectional long short-term memory (BiLSTM) optimized by variational modal decomposition (VMD) and cuckoo search (CS) algorithm. Firstly, VMD is employed to decompose the photovoltaic power sequence into sub-modes with different frequencies, and Pearson correlation analysis is adopted to determine the key meteorological factors affecting each mode. Secondly, the hybrid photovoltaic power prediction models of attention mechanism (AM) and BiLSTM are constructed, and the CS algorithm is utilized to obtain the optimal weight and threshold of the network. Finally, the prediction results of different modes are superimposed to obtain the final prediction results. The effectiveness of the proposed model is verified by predicting the output power of photovoltaic power stations in Arizona.
Abstract: Text-to-image algorithm requires high image quality and text matching. In order to improve the clarity of generated images, a generative adversarial network model is improved based on existing algorithms. Dynamic memory network, detail correction module (DCM), and text image affine combination module (ACM) are added to improve the quality of generated images. Specifically, the dynamic memory network can refine fuzzy images and select important text information storage to improve the quality of images generated in the next stage. DCM corrects details and repairs missing parts of composite images. ACM encodes original image features and reconstructs parts irrelevant to the text description. The improved model achieves two goals. On the one hand, high-quality images are generated according to given texts, with contents that are irrelevant to the texts preserved. Second, generated images do not greatly rely on the quality of initial images. Through experiments on the CUB-200-2011 bird data set, the results show that compared with previous algorithm models, the Frechet inception (FID) has been significantly improved, and the result has changed from 16.09 to 10.40, which proves that the algorithm is feasible and advanced.
Abstract: Sleep problems are becoming increasingly prominent in contemporary society. Timely detection and evaluation of sleep quality can help diagnose sleep diseases. In view of the uneven development of sleep monitoring products on the market, this study builds an online real-time sleep staging system based on dual-channel EEG signals, which uses the third-party interface brain ring to obtain EEG data, and the study combines with a CNN-BiLSTM neural network model to realize online real-time sleep staging and music regulation on the personal computer (PC). The system uses the algorithm model based on both a convolutional neural network (CNN) and a bidirectional long short-term memory (BiLSTM) neural network to automatically extract features of EEG signals. CNN can extract high-order features, and BiLSTM can capture the dependence and correlation of data before and after sleep, which makes the accuracy of sleep staging higher. The experimental results show that the proposed algorithm model achieves a staging accuracy of 92.33% and a Kappa coefficient of 0.84 in the four-classification task on the Sleep-EDF public data set. The real-time sleep staging function of the system achieves a staging accuracy of 79.17% in a self-collected sleep data staging experiment, with a Kappa coefficient of 0.70. Compared with other sleep monitoring products, this system has higher accuracy in sleep staging, diversified application scenarios, and strong real-time capability and reliability. Besides, it can regulate music for users according to the staging results to improve the sleep quality of users.
Abstract: The purpose of influence maximization is to find a small group of nodes in a network that can trigger the maximum number of remaining nodes to participate in the process of information transmission. At present, the research on the influence maximization of heterogeneous information networks usually extracts homogeneous subgraphs from the network or evaluates the influence of nodes according to the meta-path of local node structure. However, it does not consider the global features of nodes and the influence loss of the final spread range of the seed set caused by the clustering phenomenon among highly influential nodes. This study proposes an influence maximization algorithm for heterogeneous information networks based on community and structure entropy, which can effectively measure the influence of nodes locally and globally. Firstly, the local structure information and heterogeneous information of nodes in the network are retained by the construction of meta-structure to measure the local influence of nodes. Secondly, the global influence of nodes is measured by the weight ratio of the community to which the nodes belong to the whole network. Finally, the final influence of nodes is calculated, and the seed set is selected. Many experiments on real data sets indicate that the proposed algorithm is effective and efficient.
Abstract: In terms of the problems such as haze residues and color distortion in existing dehazing methods, this study takes advantage of a generative adversarial network in reconstructing image super-resolution and proposes an image dehazing algorithm based on channel attention and conditional generative adversarial network (CGAN-ECA). Specifically, the network is based on the encoder-decoder structure. The generator is designed with the multi-scale residual block (MRBlk) and efficient channel attention (ECA) to expand the receptive field, extract multi-scale features, dynamically adjust the weights of different channels, and improve the utilization rate of features. In addition, the Markovian discriminator (PatchGAN) is used to evaluate images and improve the accuracy in identifying images. At the same time, a content loss is added into the loss function to reduce pixel-level and feature-level losses of dehazing images, retain more image details, and achieve high-quality image dehazing. The test results based on the public dataset RESIDE show that compared with DCP, AOD-Net, DehazeNet, and GCANet models, the proposed model increases the peak signal to noise ratio (PSNR) and the structural similarity index (SSIM) by 36.36% and 8.80%, respectively, and color distortion and haze residue are solved. Therefore, CGAN-ECA is an effective method for image dehazing.
Abstract: PM2.5 is an important indicator for measuring the concentration of air pollutants, and monitoring and predicting its concentration can effectively protect the atmospheric environment and further reduce the harm caused by air pollution. As automatic air quality monitoring stations are constructed on a large scale, the air quality prediction model built by traditional machine learning can no longer meet the current needs. This study proposes a Gaussian-attention prediction model based on the multi-head attention mechanism and Gaussian probability estimation and utilizes the data from a monitoring station in Shenyang for training and tests. Because PM2.5 concentration is affected by other air quality data, this model uses the information alignment of hierarchical time stamps (week, day, and hour) of air quality data as input and extracts the time-series correlation features of different subspaces with the multi-head attention mechanism. More complete and effective feature information is thereby obtained, and prediction results are then acquired by Gaussian likelihood estimation. A comparison with multiple benchmark models is conducted, and the mean squared error (MSE) and mean absolute error (MAE) of the proposed Gaussian-attention prediction model are respectively 21% and 15% lower than that of the DeepAR model. Effectively improving prediction accuracy, the proposed model can accurately predict PM2.5 concentration.
Abstract: As the network expands, the exact algorithm of closeness centrality has low efficiency. In this study, a model based on the learning to rank algorithm (RankNet) is proposed to quickly approximate the closeness centrality rank of complex network nodes. Firstly, the study carries out a correlation analysis to obtain important node indicators positively correlated with the closeness centrality and put them as input features of the model. Subsequently, a subset of nodes in a given network is randomly selected and used for the training sample data of the model. The proposed model is verified by a real aviation network dataset and typical complex network models. The experimental results show that the RankNet-based model not only reduces the computational complexity but also keeps a high accuracy of the approximation. In addition, the ranking performance of the proposed model is significantly superior to that of the benchmark model based on regression learning.
Abstract: Conformance checking refers to the alignment between a computational process model and its actual execution. Conformance checking at runtime has become a new problem in current conformance checking due to the real-time feedback and positive application prospects. For each newly generated event, how to calculate and obtain the optimal conformance checking at a low performance cost is a difficult point for conformance checking at runtime. Based on the refined process structure tree (RPST) of the process model, this study proposes a conformance monitoring tree (CMT) and a dynamic programming algorithm to obtain the optimal conformance result based on the CMT. Through three experimental datasets, it is shown that compared with the existing work, the proposed algorithm has obvious performance advantages.
Abstract: As the main technical means of computer protection, intrusion detection technology has been widely studied due to its advantages of strong adaptability and ability to identify new types of attacks. However, the recognition rate and false alarm rate are difficult to guarantee, which is the main bottleneck of this technology. To improve the recognition rate and reduce the false alarm rate of anomaly detection technology, this study proposes a terminal-level intrusion detection algorithm (TL-IDA). In the data preprocessing stage, the terminal log is cut into continuous and small-block command sequences, and common statistical indicators are introduced to construct feature vectors for the command sequences. Then TL-IDA is applied to model users through the feature vectors. On this basis, a sliding window discrimination method is also proposed to judge whether the system is under attack, so as to improve the performance of the intrusion detection algorithm. The experimental results show that the average recognition rate and false alarm rate of the TL-IDA are 83% and 15%, respectively, which are superior to those of similar terminal-level intrusion detection algorithms based on anomaly technology such as ADMIT and hidden Markov model.
Abstract: Traditional image stitching algorithms are slow and fail to meet the requirements of obtaining large-resolution panoramic images in real time. To solve these problems, this study proposes an image registration algorithm based on CUDA’s speeded-up-robust features (SURF) and carries out CUDA parallel optimization on the detection and description of feature points of traditional SURF algorithms in terms of GPU thread execution model, programming model, and memory model. In addition, based on FLANN and RANSAC algorithms, the study adopts a bidirectional matching strategy to match features and improve registration accuracy. The experimental results show that compared with serial algorithms, the proposed parallel algorithm can achieve an acceleration ratio of more than 10 times for images with different resolutions, and the registration accuracy is 17% higher than that of traditional registration algorithms, with an optimal accuracy of as high as 96%. Therefore, the SURF algorithm based on CUDA acceleration can be widely used in the field of security monitoring to realize the real-time registration of panoramic images.
Abstract: It is an important step for quantum computing to outperform classical computing by evaluating quantum chips’ performance during their development to calibrate the degree of fit between the actual execution results and theoretical results of quantum algorithms. However, at present, there is no unified benchmark for evaluating the performance of quantum chips both in China and abroad, and the evaluation standards for local indicators of quantum chips can easily lead to misunderstandings about the overall performance of the chips. In view of this, this study first briefly describes performance indicators of existing quantum chips, then reviews current quantum chip evaluation methods by classifying evaluation technologies, and finally summarizes the existing problems of quantum chip evaluation technologies and looks forward to the future evaluation technology. In addition, the paper can be easily accessed by those working in the relevant fields.
Abstract: Insufficient training data is often faced in the task of text intent detection, and due to the discreteness of text data, it is difficult to perform data augmentation and improve the performance of the original model with the unchanged label. This study proposes a method combining stepwise data augmentation with a phased training strategy to solve the above problems in the few-shot intent detection. The method progressively augments the original data on whole statements and sample pairs in the same category from both global and local perspectives. During model training, the original data is learned according to different partition stages of the progressive level. Finally, experiments are performed on multiple intent detection datasets to evaluate the validity of the method. The experimental results show that the proposed method can effectively improve the accuracy and the stability of the few-shot intent detection model.
Abstract: Traditional low-power adaptive hierarchical cluster protocols in wireless sensor networks have high node energy consumption, short network lifetime, and unbalanced load. In order to solve these problems, this study proposes a Harris hawks routing optimization algorithm that reflects multi-objective cluster head election and is based on simulated annealing in heterogeneous sensor networks. On the basis of calculating the optimal threshold of nodes, the improved algorithm firstly constructs a new fitness function considering energy consumption and load to find the optimal cluster head node and ensure the uniform distribution of cluster head nodes. Then, a path selection strategy based on Harris hawks optimizer is established, and the simulated annealing algorithm is embedded to prevent from premature local optimum. Finally, the study uses an evaluation function to select cluster heads that can be added to the optimal path to shorten the communication distance between cluster head nodes and base stations. The simulation results show that compared with the CREEP, LEACH-C, and LEACH algorithms, the proposed algorithm prolong the network lifetime by 22.18%, 77.83%, and 180.52%, respectively, and thus they can prolong the network lifetime more effectively.
Abstract: When designing a fault-tolerant scheduling algorithm in a real-time heterogeneous system, it is necessary to consider the constraints of real-time and maximize the reliability of the system. Furthermore, parallel application scheduling problems in heterogeneous systems have been shown to be NP-complete. Most of the existing fault-tolerant scheduling algorithms use replication technology to improve the reliability of the system, but the multiple execution of tasks will lead to longer application execution time and reduced system real-time performance. Therefore, a fault-tolerant scheduling algorithm based on active replication technology is proposed. The algorithm continuously replicates the tasks that have the least impact on the real-time performance of the current system in the task set and then schedules all tasks in the task set to the earliest completed processor. While meeting the real-time constraints, the algorithm improves the reliability of the system. Experiments show that the reliability of the proposed algorithm has been improved under strict time constraints compared with that of the DB-FTSA algorithm which also focuses on real-time heterogeneous systems.
Abstract: Individual electroencephalogram (EEG) signals are different and vulnerable to environmental factors. In view of these problems, this paper adopts methods of removing baseline interference and EEG channel selection and proposes an emotion classification and recognition algorithm based on a continuous convolutional neural network (CNN). Firstly, the selection of differential entropy (DE) characteristics of baseline signals is studied. After the data is processed into multi-channel input, the continuous CNN is used for classification experiments, and then the optimal number of electrodes is determined. The experimental results show that after the difference between the DE of experimental EEG signals and that of baseline signals of the subject one second before the experimental EEG is mapped into a two-dimensional matrix, and the frequency dimension is turned into a multi-channel form to serve as the input of the continuous CNN, the average classification accuracy of arousal and valence on 22 channel is 95.63% and 95.13%, respectively, which are close to that on 32 channel.
Abstract: While ensemble learning has achieved remarkable success in generalization performance, the error analysis of ensemble learning needs further research. As cross-validation has an important application for model performance evaluation in statistical machine learning, block-3×2 cross-validation and k-fold cross-validation are applied to integrate the weighted prediction values for each sample point and analyze the error. Experiments on simulated data and real data show that the prediction error of ensemble learning based on block-3×2 cross-validation is smaller than that of a single learner, and the variance of ensemble learning is smaller than that of a single learner. The generalization error of the ensemble learning based on block-3×2 cross-validation is less than that of the one based on k-fold cross-validation, which indicates that the ensemble learning model based on block-3×2 cross-validation has good stability.
Abstract: Football match scenes are featured with dense crowds and many mobile targets, and YOLOv3 algorithm has low detection accuracy and requires massive model parameters, which makes it unable to be deployed on mobile devices with limited computing power. In view of these problems, this study proposes a pedestrian detection method based on improved YOLOv3. Specifically, the study replaces the Darknet-53 backbone feature extraction network with a more efficient and lightweight GhostNet network, selects detection branch layers with four scales, and adopts the K-means++ algorithm to improve the clustering effect of the anchor box. Furthermore, the study adds spatial pyramid pooling to achieve an output with the same size as the input image, puts forward the CIoU loss function to calculate the loss value of target positioning, adds heatmap visualization, and uses Mosaic data enhancement in training. The experimental results show that YOLOv3-GhostNet achieves a mAP of 90.97% on the VOC fusion dataset, with an improvement of 1.75% compared with the YOLOv3 algorithm. In addition, it reduces the number of parameters by about 81.4% and increases the real-time detection rate by about 1.5 times, which shows a positive detection effect on small mobile devices.
Abstract: Trucks cannot be accurately identified when they do not follow the prescribed time and route on urban roads by avoiding cameras and other means. In view of this, an urban road truck detection method based on improved Faster RCNN is proposed. Features are extracted by performing convolution and pooling operations on the vehicle images passed into the backbone network. The feature pyramid network (FPN) is added to improve the accuracy of multi-scale target detection. At the same time, the K-means clustering algorithm is applied to the dataset to obtain new anchor boxes. Region proposal network (RPN) is utilized to generate proposal boxes and complete-IoU (CIoU) loss function is used for replacing the smoothL1 loss function of the original algorithm to improve the accuracy of vehicle detection. The experimental results show that the improved Faster RCNN increases the average precision (AP) for truck detection by 7.2% and the recall by 6.1%. The improved method reduces the possibility of missed detection and has a good detection effect in different scenarios.
Abstract: Assisting users in understanding the clauses of insurance products is one of the hot issues in insurance applications. It is feasible to assist the life insurance business with knowledge graph technology. The life insurance knowledge graph (LIKG) is extracted and constructed by multi-source data. Specifically, the BERT-IDCNN-BiLSTM-CRF model is applied to extract entities from unstructured data, and the entity is aligned by a variety of short text similarity algorithms and ranking ensemble algorithm. A two-stage extraction algorithm is designed to fill the attributes of insurance products by Bootstrapping and classification prediction. Then a prototype system is designed based on LIKG. The system uses the entity extraction and the attribute extraction to provide knowledge acquisition, designs an index called CF-IIF to provide attribute recommendation function, and realizes a visual interface to help users quickly master the information of life insurance, which demonstrates the application value of LIKG.
Abstract: Given the various problems of the cultural algorithm, such as slow convergence speed, high likeliness to fall into local optimum, and low population diversity, this study optimizes the design of the cultural algorithm and proposes a hybrid optimization algorithm that incorporates a genetic algorithm (GA) with an elite retention strategy and a simulated annealing (SA) algorithm into the framework of the cultural algorithm (CA). In light of the idea of co-evolution, this algorithm is divided into a lower population space and an upper belief space that share the same evolutionary mechanism but use different parameters. On the basis of the CA, a GA with an elite retention strategy is added so that the outstanding individuals in the population can directly enter the next generation to improve the convergence speed. An SA algorithm is added as its mutation characteristics can be leveraged to enable the algorithm to probabilistically jump out of the local optimum and accept inferior solutions and thereby increase population diversity. The function optimization results prove the effectiveness of the proposed algorithm. This algorithm is applied to solve the flow shop scheduling problem of minimizing the maximum completion time. The simulation results show that the proposed algorithm is superior to several other representative algorithms in convergence speed and accuracy.
Abstract: Considering that shadows caused by changes in lighting are difficult to identify and segment for intelligent surveillance videos in indoor environments, this study proposes a UNet network combining the transfer learning method and the SENet channel attention mechanism. Specifically, because shadow features are blurry and difficult to extract effectively, the SENet channel attention mechanism is added to the upsampling part of the UNet model to improve the feature weight of the effective area without increasing the network parameters. A pre-trained VGG16 network is then migrated into the UNet model to achieve feature migration and parameter sharing, improve the generalization ability of the model, and reduce training costs. Finally, the segmentation result is obtained by a decoder. The experimental results show that compared with the original UNet algorithm, the improved UNet algorithm offers significantly enhanced performance indicators, with its segmentation accuracy on moving objects and shadows respectively reaching 96.09% and 92.24% and a mean intersection-over-union (MIOU) of 92.58%.
Abstract: The core of server cache performance is the cache replacement strategy which directly affects the cache hit ratio. Web cache can solve the problems of network congestion and user access delay and improve server performance. A multi-cache replacement strategy based on spectral clustering is proposed because of the low cache hit ratio of traditional cache replacement algorithms. The strategy uses the circular sliding window mechanism to extract multiple temporal features and access attributes of log files and conducts cluster analysis on the filtered data set through spectral clustering to obtain access prediction results. Multi-cache replacement strategy takes into account the local frequency, global frequency, and resource size of the cache object to eliminate the low-value resources and retain the high-value resources. In comparison with traditional replacement algorithms such as LRU, LFU, RC, and FIFO, the experimental results show that the combination of spectral clustering and multi-cache replacement strategy in this paper can effectively improve the cache request hit ratio and byte hit ratio.
Abstract: In order to quickly drive accident vehicles away from the scene and ensure a clear road during a minor traffic accident, this study proposes a vehicle collision detection and liability determination model. First, the study combines the SSD (single shot multibox detector) target detection algorithm and the MobileNet lightweight deep network model to make improvements and obtain the position and size information of the moving target in each frame of video images, so as to identify and detect the vehicle. Secondly, the study employs a Kalman filter to establish a corresponding matching relationship between moving targets in consecutive image frames, predict their motion states, and judge their positions and motion trend, in a bid to track the vehicle. Then, the study determines whether there is a collision by the intersection over union of the vehicle target detection frame. Finally, according to the speed and direction information of the vehicle on a straight road, the liability of the accident vehicle is determined under the road safety regulations and the fast method of motor vehicle accidents. The results show that the research can help to detect and determine the liability during vehicle collisions caused by rear-end collisions and lane changes on straight roads.
Abstract: The existing news recommendation system fails to sufficiently consider the semantic information of news, and modeling factors for news body suffers from unity problems. Attention-BodyTitleEvent (Attention-BTE), a news recommendation algorithm based on fusion of attention and multi-perspectives, is proposed in this study. The BERT model and attention mechanism are applied to vectorize the body, title, and event in the news respectively. The three parts are combined to represent news vectorization, and then the candidate news and user browsing news data are processed respectively to obtain the corresponding candidate news vectorization and user vectorization. Finally, dot multiplication is conducted to obtain the probability of users clicking on the candidate news, namely the news recommendation result. Experimental data demonstrate that Attention-BTE improves the index by about 6% compared with the other news recommendation algorithm.
Abstract: The object detection algorithms based on the feature pyramid network do not give due consideration to the scale differences among different objects and the high-frequency information loss during cross-layer feature fusion, denying the network sufficient fusion of global multi-scale information and consequently resulting in poor detection effects. To solve these problems, this study proposes a scale-enhanced feature pyramid network. This method improves the lateral connection and cross-layer feature fusion modes of the feature pyramid network. Specifically, a multi-scale convolution group with the dynamic receptive field is designed to serve as a lateral connection so that the feature information of each object can be extracted sufficiently, and a high-frequency information enhancement module based on the attention mechanism is introduced to promote the fusion of high-layer features with low-layer ones. The experimental results on the MS COCO dataset show that the proposed method can effectively improve the detection accuracy on objects at each scale and its overall performance is better than that of the existing methods.
Abstract: In the cloud storage environment, data owners can store and share data through cloud servers, but the following security issues may exist. First, data owners need to guarantee the authentication of their data. Secondly, the data may contain the data owner’s sensitive information, such as name, age, and other information. Therefore, data owners may reveal their sensitive information when sharing data with other users. To solve the above problems, this study proposes a certificateless sanitizable signature scheme to ensure the authentication of shared data and the sensitive information hiding in cloud storage environments. Specifically, the proposed scheme is based on certificateless cryptography, which avoids the high certificate management overhead in traditional public key infrastructure and eliminates the key escrow defect in identity-based cryptography. In addition, the scheme adds access control, so that the data stored in the cloud server can only be accessed by authorized users. Finally, the security analysis shows the security of the scheme and the performance analysis reflects the efficiency of the scheme.
Abstract: As information technology develops, recommendation system serves as an important tool in the era of information overload and plays an increasingly important role. Traditional recommendation systems based on content and collaborative filtering tend to model the interaction between users and items in a static way to obtain users’ previous long-term preferences. Because users’ preferences are often dynamic, unsustainable, and behavior-dependent, sequential recommendation methods model the interaction histories between users and items as ordered sequences, which can effectively capture the dependencies between items and users’ short-term preferences. However, most sequential recommendation models overemphasize the behavior order of user-item interaction and ignore the temporal information in interaction sequences. In other words, they implicitly assume that adjacent items in the sequences have the same time interval, which leads to limitations in capturing users’ preferences that include temporal dynamics. In response to the above problems, this study proposes a self-attention-based network for time-aware sequential recommendation (SNTSR) model, which integrates temporal information into an improved self-attention network to explore the impact of dynamic time on the prediction of the next item. At the same time, SNTSR independently calculates position correlation to eliminate the noise correlations that may be introduced and enhance the ability to capture users’ sequential patterns. Extensive experimental studies are carried out on two real-world datasets, and results show that SNTSR consistently outperforms a set of state-of-the-art sequential recommendation models.
Abstract: The generation of text adversarial samples is of great significance for studying the vulnerability of deep learning-based natural language processing (NLP) systems and improving the robustness of such systems. This work studies the important steps in the generation of word-level adversarial samples and the search for replacement words. Considering the problems of premature convergence and poor effectiveness of existing algorithms, a text adversarial sample generation method is proposed, which is based on an improved artificial bee colony (ABC) search algorithm. Firstly, the search space of the words to be replaced is obtained by the screening of the sememe annotations of the words in the HowNet database. Then, the improved ABC algorithm is employed to search and locate the replacement words for the generation of high-quality text adversarial samples. Finally, attack tests are conducted on two text classification datasets for a comparison with the current mainstream text classification models based on deep neural networks (DNNs). The results demonstrate that compared with the existing text adversarial sample generation methods, the proposed method can mislead the text classification system with a higher success rate of attack and preserve semantic and grammatical correctness to a larger extent.
Abstract: Cataract is an ocular disease that mainly causes visual impairment and blindness, and early intervention and cataract surgery are the primary ways of improving the vision and the life quality of cataract patients. Anterior segment optical coherence tomography (AS-OCT) is a new type of ophthalmic image featuring non-contact, high resolution, and quick examination. In clinical practice, ophthalmologists have gradually used AS-OCT images to diagnose ophthalmic diseases such as glaucoma. However, none of the previous works have focused on automatic cortical cataract (CC) classification with such images. For this reason, this study proposes an automatic CC classification framework based on AS-OCT images, and it is composed of image preprocessing, feature extraction, feature screening, and classification. First, the reflective region removal and contrast enhancement methods are employed for image preprocessing. Next, 22 features are extracted from the cortical region by the gray level co-occurrence matrix (GLCM), grey level size zone matrix (GLSZM), and neighborhood gray-tone difference matrix (NGTDM) methods. Then, the Spearman correlation coefficient method is used to analyze the importance of the extracted features and screen out redundant ones. Finally, the linear support vector machine (linear-SVM) method is utilized for classification. The experimental results on a clinical AS-OCT image dataset show that the proposed CC classification framework achieves 86.04% accuracy, an 86.18% recall rate, 88.27% precision, and 86.35% F1-score respectively and obtains performance comparable to that of the advanced deep learning-based algorithm, indicating that it has the potential to be used as a tool to assist ophthalmologists in clinical CC diagnosis.
Abstract: The promotion of digital braille in the information age can help to improve the cultural quality and living standards of blind people in China. This study implements a Chinese-braille conversion system based on the national general braille (NGB) tone rules, which can quickly generate a large number of digital resources in line with the NGB rules and make visually impaired people obtain information without barriers. This system processes Chinese text according to the NGB tone rules and converts it into braille that conforms to tone rules and abbreviation rules. The test results show that the system can accurately process the tone rules and abbreviation rules and obtain accurate digital braille that is in line with the NGB tone rules. In addition, the coverage rate of tonal and final abbreviations and the increase in length are all comparable to the theoretical values of the NGB tone rules. The system can quickly process long-form corpus files and execute programs efficiently. Furthermore, it has practical value and can be used to promote the NGB and promote the barrier-free construction of digital braille in China.
Abstract: Block diagonalization (BD) belongs to a traditional linear precoding algorithm with multiple inputs and outputs, and its core idea is to find the orthogonal basis of the null space in interference matrixes through singular value decomposition (SVD), so as to eliminate the multiuser interference (MUI). However, as the number of transmitters and receivers increases, the BD precoding algorithm faces more complex computation, which has become one of the key factors restricting its development. Therefore, this study proposes an optimal low-complexity BD algorithm. The algorithm is based on the combination algorithm of Schmidt orthogonalization inversion and lattice reduction operation in orthogonal decomposition, and it replaces the SVD of two high-complexity operations on the traditional BD algorithm by Schmidt orthogonalization inversion and lattice reduction operation and thus reduces the algorithm complexity. The results show that the computational complexity of the optimal algorithm is reduced by 46.7%, and the system and capacity are increased by 2–10 bits/Hz. Furthermore, the bit error rate is improved by two orders of magnitude.
Abstract: Amid the further development of the marine meteorological business, marine meteorological services are gradually developing towards specialization, visualization, and intelligence. As a result, comprehensive marine meteorological services can no longer meet the actual business needs of meteorological services to ports. To ensure the safety of port production and improve the efficiency of meteorological services to ports, this study proposes a construction scheme for an intelligent meteorological service system for ports based on the service-oriented architecture (SOA). Multi-source heterogeneous business data, such as meteorological, port, and geographic information, are dynamically integrated, and extensive markup language (XML), Web service, data warehouse, middleware mode, WebGIS, message queue, and other computer-related technologies are employed. Various functions are thereby fulfilled, including real-time monitoring of meteorological business data of a port area, professional forecast and early warning for ports, preparation and release of emergency plans, and threshold management of professional users and meteorological elements. The business application results of the proposed system show that the system deserves application and promotion as it meets the demand of professional meteorological services to ports, effectively reduces the adverse impact of marine meteorological disasters on the production activities in the port area, and is highly scalable.
Abstract: Program dependency graph usually judges the data dependency according to definition-use relationships of variables in statements, and it cannot make an accurate judgment according to the semantics, which leads to the introduction of false dependency relationships and the repair failure caused by the use of error information in repairing defects. Therefore, this study will prune false dependencies related to null objects or null pointers by using abstract attributes and propose an abstract semantic-based program dependency graph to reduce the analysis of dependency relationships unrelated to the semantics of program defects and repair null pointer references. Based on the dependency relationships obtained from the analysis, a multi-strategies repair scheme is implemented under the guidance of different repair strategies for null pointer references, and the null pointer references are repaired with side effects minimized as much as possible. In addition, in this study, the null pointer references in Defects4J are adopted to evaluate the repair tool DTSFix through experiments. The results show that the repair effect of DTSFix is much better than that of other tools, which proves the effectiveness of the method.
Abstract: Nowadays, the global navigation satellite system (GNSS) has basically achieved real-time positioning with high precision in outdoor open environments. With the acceleration of urbanization, however, providing pedestrian navigation services for densely built-up sites disturbed by GNSS signals generates great demand, which has significantly promoted indoor positioning technology in recent years. Furthermore, as there is no single universal positioning method to realize the seamless transition between indoor and outdoor environments, seamless navigation technology introduces new hot spots and research topics to solve the “last-kilometer” problem in the navigation field. This study summarizes the multi-sensor fusion technology for indoor pedestrian navigation: (1) The advantages and limitations of a single sensor in indoor positioning are analyzed and compared from the perspective of radio frequency signals and non-electrical signals separately; (2) the positioning methods in the field of indoor multi-sensor fusion are introduced, including multimodal fingerprint fusion, geometric ranging fusion, and PDR-based fusion. Finally, the solution to the application of indoor positioning technology in seamless navigation is studied, and the challenges and prospects of seamless positioning in indoor and outdoor environments are presented. The research provides references and assistance to the follow-up research on high-precision seamless positioning.
Abstract: In this study, a multimodal emotion recognition method is proposed, which combines the emotion recognition results of speech, electroencephalogram (EEG), and faces to comprehensively judge people’s emotions from multiple angles and effectively solve the problems of low accuracy and poor robustness of the model in the past research. For speech signals, a lightweight fully convolutional neural network is designed, which can learn the emotional characteristics of speech well and is overwhelming at the lightweight level. For EEG signals, a tree-structured LSTM model is proposed, which can comprehensively learn the emotional characteristics of each stage. For face signals, GhostNet is used for feature learning, and the structure of GhostNet is improved to greatly promote its performance. In addition, an optimal weight distribution algorithm is designed to search for the reliability of modal recognition results for decision-level fusion and thus more comprehensive and accurate results. The above methods can achieve the accuracy of 94.36% and 98.27% on EMO-DB and CK+ datasets, respectively, and the proposed fusion method can achieve the accuracy of 90.25% and 89.33% on the MAHNOB-HCI database regarding arousal and valence, respectively. The experimental results reveal that the multimodal emotion recognition method proposed in this study effectively improves the recognition accuracy compared with the single mode and the traditional fusion methods.
Abstract: As the terrain generation algorithm has trouble balancing ease of use, controllability, realism, and speed, this study proposes a terrain generation method based on sketch maps. This method uses the generative adversarial network to model the terrain slope, slope aspect, and other information in the hidden space so that the generated terrain conforms to the constraints of the user’s hand-drawn sketch map. This study also proposes a sketch map extraction algorithm based on terrain height maps, and it can extract a sketch map to a hand-drawn effect from a real terrain height map and quickly build the data set. An algorithm for multi-scale terrain detail filling is proposed. Owing to the introduction of the multi-scale concept, the terrain texture details are dynamically supplemented, and the realism and aesthetic properties are greatly improved. A terrain satisfaction evaluation method based on user feedback is proposed and verified by experiments. The results show that the proposed terrain generation method can accurately and efficiently generate digital terrain that meets users’ expectations.
Abstract: The current deep learning models in the field of compilation optimization generally perform single-task learning and fail to use the correlation among multiple tasks to improve their overall compilation acceleration effect. For this reason, a compilation optimization method based on multi-task deep learning is proposed. This method uses the graph neural network (GNN) to learn program features from the abstract syntax trees (ASTs) and control data flow graphs (CDFGs) of the C program and then predicts the initiation interval and loop unrolling factor for the software pipelining of the HX digital signal processor (HXDSP) synchronously according to program features. Experimental results on the DSPStone dataset show that the proposed multi-task method achieves a performance improvement of 12% compared with that of the single-task method.
Abstract: In federated learning, due to barriers such as industry competition and privacy protection, users keep data locally and cannot train models in a centralized manner. Users can train models cooperatively through the central server to fully utilize their data and computing power, and they can share the common model obtained by training. However, the common model produces the same output for different users, so it cannot be readily applied to the common situation where users’ data are heterogeneous. To solve this problem, this study proposes a new algorithm based on the meta-learning method Reptile to learn personalized federated learning models for users. Reptile can learn the initial parameters of models efficiently for multi-tasks. When a new task arrives, only a few steps of gradient descent are needed for convergence to satisfactory model parameters. This advantage is leveraged, and Reptile is combined with federated averaging (FedAvg). The user terminal uses Reptile to process multi-tasks and update parameters. After that, the central server performs the averaging aggregation of the parameters the user updates and iteratively learns better initial parameters of the model. Finally, after the proposed algorithm is applied to each user’s data, personalized models can be obtained by only a few steps of gradient descent. In the experiment, this study uses simulated data and real data to set up federated learning scenarios. The experiment shows that the proposed algorithm can converge faster and offer a better personalized learning ability than other algorithms.
Abstract: The variable scales of objects and the use of feature fusion have been the challenges for popular object detection algorithms. Considering the problems, this study proposes a multi-path feature fusion module, which strengthens the connection between input and output features and alleviates the dilution of feature information in transmission by adopting cross-scale and cross-path feature fusion. Meanwhile, the study also proposes a scale-aware module by refining the attention model, which allows the model to easily recognize multi-scale objects by selecting the size of the receptive field corresponding to the scale of the objects independently. After the scale-aware module is embedded into the multi-path feature fusion module, the feature extraction and utilization abilities of the model are improved. The experimental results reveal that the proposed method achieves 82.2 mAP and 38.0 AP on PASCAL VOC and MS COCO datasets, respectively, an improvement of 1.3 mAP and 0.6 AP over the baseline FPN Faster RCNN, respectively, with the most significant improvement in detection of small-scale objects.
Abstract: Object images in the real world often have large intra-class variations, and thus using a single prototype to describe an entire category will lead to semantic ambiguity. Considering this, a multi-prototype generation module based on superpixels is proposed, which uses multiple prototypes to represent different semantic regions of objects and employs the context to correct prototypes among the generated prototypes by a graph neural network to ensure the orthogonality of the sub-prototypes. To obtain a more accurate prototype representation, a Transformer-based semantic alignment module is designed to mine the semantic information contained in the features of the query images and the background features of the supporting images. In addition, a multi-scale feature fusion structure is proposed to instruct the model to focus on features that appear in both the supporting images and the query images, which can improve the robustness to changes in object scales. The proposed model is tested on the PASCAL-5i dataset, and the mean intersection over union (mIoU) is improved by 6% compared with that of the baseline model.
Abstract: The network information-centric system of systems (SoS) is a new generation of command-and-control operational SoS proposed by the PLA, which has the advantage of dynamic response to missions and environmental changes. It optimizes the operational resources of the whole network to maximize operational effectiveness. With the development of artificial intelligence and other technologies, the current optimization method, which mainly depends on the implementation of plans, can neither adapt to the self-evolution of intelligent and unmanned equipment nor cover the battlefield dynamics. Considering the above defects, this study takes the air and missile defense operational SoS as an example to study the solution to the resource integration scheme in the case of physical node damage. The down-selection model is adopted to transform the solution to the resource integration scheme into a combinatorial optimization problem, and the formation mechanism of initial evolutionary strategy is improved by adding disturbance restrictions. Thus, a resource optimization method based on the evolutionary game is proposed. The effectiveness of the method is verified by simulations on the Netlogo platform. Compared with the result of the resource optimization method based on the genetic algorithm, the task completion of the solution by the proposed method is increased by 6.4% on average.
Abstract: A human-machine collaborative classification model based on the deep forest is proposed to solve the problems encountered in the classification of defects in industrial products, such as sample image shortage, insufficient classification precision, and time-consuming model training. For this purpose, sample images are preliminarily identified with deep forest, and their features are extracted by the multi-granularity scanning module and the cascaded forest module. The initial estimation result is thereby obtained, and sample images difficult to identify are separated out. Then, the human-machine collaboration strategy is employed. Specifically, some of the sample images difficult to identify are randomly labeled manually, and the remaining ones are reclassified with the K-nearest neighbor algorithm. The experimental results on the public dataset and the real data collected from the production line indicate that the improved classification model offers performance superior to that of the baseline algorithm on the dataset of industrial product surface defects.
Abstract: The spatial-temporal distribution of crowds as well as evacuation safety and efficiency are impacted by the layout of indoor obstacles. To investigate the effect, a crowd evacuation model for a single room with a single exit and obstacles is proposed in this study. Three different influencing factors (i.e., obstacle length, the distance between an obstacle and the exit, and the distance between the obstacle center and exit center) are used to analyze their impacts on evacuation efficiency and safety. The experimental results reveal that the obstacle length is directly proportional to evacuation efficiency and is inversely proportional to evacuation safety. The evacuation efficiency and safety are directly proportional to the distance between the obstacle and the exit and are inversely proportional to the distance between the obstacle center and exit center. Additionally, a multi-objective evolutionary algorithm is used to optimize the layout of the indoor obstacles, and the obtained results can provide an important reference for decision-makers to balance evacuation efficiency and safety.
Abstract: Intelligent disinfection robots are a highly effective way of daily disinfection as it becomes regular. Robots usually perceive the surrounding environment through vision, but object detection based on supervised learning usually requires a large amount of labeled data for training. When the amount of labeled data is large, the cost of labeling is very high, and when the amount of labeled data is small, the model is prone to overfitting. Therefore, few-shot object detection is an effective solution. On the basis of the SimDet Model, this study proposes the SimDet+ model. First, according to the characteristics of the object detection task in a disinfection scene, the process of self-supervised pre-training is added. Second, as there are query images for reference, the classification layer is improved, where the cosine similarity instead of the fully connected layer is employed for confidence level calculation, and thus the overfitting phenomenon is effectively avoided through non-parametric calculation. For the disinfection scene, a 22-minute video dataset and a detection dataset containing eight categories of objects are produced and used in two stages separately for training. Through self-supervised pre-training, the cost of data labeling is effectively reduced, and the mAP of downstream tasks is increased from 0.216 2 to 0.530 2.
Abstract: A lightweight object detection network YOLOv5-tiny is given on the basis of YOLOv5 for real-time target detection tasks in drone-captured scenarios. The replacement of the original backbone network CSPDarknet53 with MobileNetv3 reduces the parameters of the network model and substantially improves the detection speed. Furthermore, the detection accuracy is improved by the introduction of the CBAM attention module and the SiLU activation function. With the characteristics of the aerial photography task dataset VisDrone, the anchor size is optimized, and data augmentation methods such as Mosaic and Gaussian blur are used to further improve the detection effect. Compared with the results of the YOLOv5-large network, the detection efficiency (FPS) is improved by 148% at the expense of a 17.4% reduction in mAP. Moreover, the network size is only 60% of that of YOLOv5 when the detection results are slightly superior.
Abstract: With the improvement of urban residents’ consciousness of green and low-carbon travel, ride-sharing travel of online car-hailing emerged at the historic moment. However, due to the driving route issue involved in the ride-sharing mode, differences between the passengers and between the passengers and the driver are highly likely to occur. Moreover, the costs of this ride-sharing travel mode remain to be clarified. For the above-mentioned and various other reasons, the ride-sharing mode has not been widely promoted and applied. To address the problems with this travel mode, this study constructs a route optimization model for online car-hailing ride-sharing. and the model takes into account the cost of waiting time, that of driving distance, benefit, capacity constraint, time window constraint, etc. According to the characteristics of the ride-sharing model, a solving genetic algorithm satisfying the constraints on the ride-sharing model is designed by drawing on the genetic algorithm. Matlab software is used to run the algorithm program and thereby solve the calculation example, and a maximum profit of 6 906.297 1 CNY and the detailed driving route for the vehicle are obtained after the program is run 44.08 s. The experiment shows that an approximate optimal solution for the ride-sharing route can be obtained by the ride-sharing model of online car-hailing constructed and the genetic algorithm designed, which proves the feasibility and effectiveness of the proposed model and algorithm.
Abstract: To study the electric field effect of conductive droplets with low conductivity, an electrohydrodynamic atomization (EHDA) solver based on the leaky dielectric model and the volume of fluid (VOF) method is designed by the computational fluid dynamics (CFD) software OpenFOAM. The numerical results are compared with Taylor’s analytical values, and the simulation results predict the deformation ways of droplets and the mode of circumfluence inside and outside the droplets. It is found that under the action of an external electric field, the droplets will become “prolate” or “oblate” and form stable circumfluence inside, and they only undergo deformation without any macroscopic motion. As the intensity of the electric field increases, the deformation of the droplets also intensifies. In the case of small deformation, the simulated values are consistent with the analytical values, which verifies the correctness of the numerical method. When the droplet deformation is considerable, the simulation results start to deviate from the theoretical values, which is consistent with the experimental observations. In addition, the effect of the change in conductivity on droplet deformation is also apparent, while the evolution of the dielectric constant ratio has a less pronounced impact on droplet deformation.
Abstract: In order to solve the problem of low accuracy caused by large classification loss in the lightweight target detection algorithm, a method of detecting the location and classification of the target with double detection heads is proposed. In the algorithm, the convolution head is used to detect the position, and the full connector is used to detect the classification. In the classification detection, after the feature map passes through the convolution layer, the feature map of the fused position regression branch is processed through the full connection layer. A grouping full connection method is proposed to further reduce the amount of calculation in the full connection layer. The algorithm is trained in VOC datasets. The results show that the classification loss of the improved model is significantly reduced, and the detection accuracy of the lightweight target detection algorithm is effectively improved. The accuracy of the algorithm on the VOC test set has reached 70.08% mAP.
Abstract: In the problem of predicting disease scores amid the protection of multi-source domain data considering user privacy, the decentralized data from different source domains cannot be combined and may follow different distributions. Therefore, traditional machine learning methods cannot be applied directly to utilize the information within source domains. In this study, the federated importance weighting method is proposed combining the idea of federated learning and the sample-based transfer learning approach. By re-weighting the samples from the source domains to the prediction task of the target domain, and without data sharing between multiple source domains, it realizes the use of data with different distributions while protecting the data privacy of the source domains. Moreover, this study constructs a weighted model and provides a concise and general algorithm to solve the prediction model for the target domain. Numerical simulation and empirical results show that, compared with the traditional method without considering distribution shift, the federated importance weighting method can effectively utilize the information of the source domain data. It is superior in prediction accuracy of the target domain and can make an accurate prediction of disease scores in the Parkinson’s disease data.
Abstract: Recycling and remanufacturing industrial products is conducive to reducing production costs and protecting the environment. It is very important to make excellent equipment disassembly sequence planning to improve disassembly efficiency and reduce recovery costs. For the factors of equipment recycling in actual disassembly, a disassembly sequence planning model based on a discrete whale optimization algorithm (DWOA) is proposed in this study. The objective function of the model employs the position change cost as the new evaluation indicator and adopts the stratified combination method to rapidly generate the initial population. DWOA features the precedence preservative crossover mechanism, heuristic mutation, and better global and local search ability. Comparative experiments are conducted with recycled upper rubber plate and standstill seal to test the feasibility of the proposed algorithm. The experimental results show that at the same time, the algorithm stability, optimization ability, and convergence speed of DWOA are better than those of other algorithms.
Abstract: Neural process (NP) combines the advantages of neural networks and Gaussian processes to estimate uncertainty distribution functions from a small number of contexts and implement function regression. It has been applied to a variety of machine learning tasks such as data complementation and classification. However, for 2D data regression problems (e.g., image data completion), the prediction accuracy of NP and the fitting of the contexts are deficient. To this end, an image-faced neural process (IFNP) is constructed by integrating a convolutional neural network (CNN) into the neural process based on the lower bound of evidence and loss function derivation. Then, a local pooled attention (LPA) module and a global cross-attention (GCA) module are designed for the IFNP, and an image-faced attentive neural process (IFANP) model with significantly better performance than the NP and IFNP is constructed. Finally, these models are applied to MNIST and CelebA datasets, and the scalability of IFNP is demonstrated by combining qualitative and quantitative analysis. In addition, the better data completion and detail-fitting ability of IFNP are confirmed.
Abstract: In this study, an improved model based on you only look once version 5 (YOLOv5) is proposed to solve the problem of difficult detection of small targets in images. In the backbone network, a convolutional block attention module (CBAM) is added to enhance the network feature extraction ability. As for the neck network, the bi-directional feature pyramid network (BiFPN) structure is used to replace the path aggregation network (PANet) structure and thereby strengthen the utilization of low-level features. Regarding the detection head, a high-resolution detection head is added to improve the ability of small target detection. A number of comparative experiments are conducted, respectively, on a facial blemish dataset and an unmanned aerial vehicle (UAV) dataset VisDrone2019. The results show that the proposed algorithm can effectively detect small targets.
Abstract: In many practical applications of data mining, instances for each cluster are often required to be balanced in number. However, the entropy-weighted K-means algorithm (EWKM) for independent subspace clustering leads to unbalanced partitioning and poor clustering quality. Therefore, this study defines a multi-objective entropy that takes balanced partitioning and feature distribution into account and then employs the entropy to improve the objective function of the EWKM algorithm. Furthermore, the study designs the solution process by using the iterative method and alternating direction method of multipliers and proposes the entropy-based balanced subspace K-means algorithm (EBSKM). Finally, the clustering experiments are conducted in public datasets such as UCI and UCR, and the results show that the proposed algorithm outperforms similar algorithms in terms of accuracy and balance.
Abstract: Considering the insufficient feature extraction in hyperspectral remote sensing image classification under limited training samples, a multi-scale 3D capsule network is proposed to improve hyperspectral image classification. Compared with the traditional convolutional neural network, the proposed network is equivariant, and its input and output forms are neurons in the form of vectors rather than scalar values in the convolutional neural network. It is conducive to obtaining the spatial relationship between objects and the correlation between features and can avoid problems such as overfitting under limited training samples. Specifically, the network extracts the features of an input image through the convolution kernel operation on three scales to obtain the features of different scales. Then, the three branches are connected to different 3D capsule networks to obtain the correlation between spatial spectrum features. Finally, the results of the three branches are fused, and the classification results are obtained by the local connection and margin loss function. The experimental results reveal that this method has good generalization performance on the open-source hyperspectral remote sensing data set and has higher classification accuracy than other advanced hyperspectral remote sensing image classification methods.
Abstract: In recent years, digital signal modulation recognition has gradually become an important line of research in the field of wireless communication owing to its high information confidentiality and anti-noise ability. As one of the important features of modulation recognition, the constellation diagram has obvious advantages in feature extraction because it does not need to receive prior information on the signal during feature extraction. For the above reasons, this study presents an overview of digital signal modulation methods based on the constellation diagram. In particular, it starts by analyzing the basic principle of the constellation diagram. Then, the characteristics of the constellation diagram in various lines of research are analyzed by summarizing the existing digital signal modulation recognition schemes based on the constellation diagram. Finally, the development trend and future expectations of digital modulation recognition schemes based on the constellation diagram are presented.
Abstract: Air pollution is an important factor affecting public health, and air quality prediction is the key to air pollution early warning and a hot research topic in the fields of environmental science, statistics, and computer science in recent years. This study reviews the research status and progress of air quality prediction methods, with a special focus on the systematical analysis and summarization of the applications of the newly-emerged deep learning methods in recent years in air quality prediction. Specifically, the evolution process of air quality prediction methods and air pollution datasets are outlined. After the traditional air quality prediction methods are described, the progress of existing deep learning-based air quality prediction methods is analyzed and compared in detail from the perspectives of temporal information, temporal-spatial information, and attention mechanisms. Finally, the development trend of air quality prediction methods is summarized and predicted.
Abstract: In vehicular named data network (VNDN), interest flooding attack (IFA) occupies or even exhausts network resources by sending a large number of malicious interest packets, which results in the failure to meet the requests of legitimate users and seriously endangers the operation safety of Internet of Vehicles (IoV). To solve the above problems, this study proposes an IFA detection method based on traffic monitoring. Firstly, a distributed network traffic monitoring layer based on RSU is constructed, where each RSU monitors the network traffic within its communication range, and the communication interconnection between RSUs forms the RSU network traffic monitoring layer. Secondly, a fixed time window is set, and the network traffic in each window is analyzed from three dimensions, i.e., information entropy, network self-similarity, and singularity. Additionally, a new field is added to the interest packet, and thus information entropy can be used to reflect the distribution of interest packet sources. Finally, the above three indicators are comprehensively employed to judge the existence of attack. The simulation results indicate that the proposed method effectively improves the accuracy of IFA detection and reduces the misjudgment rate.
Abstract: With the rapid increase in the number of Android applications, more importance is attached to the quality of Android applications. Testing is an important guarantee for high-quality software, and test case generation technology is the key to automated testing. Data shows that nearly 88% of Android applications in Google Play use reflection. The existing automatic test case generation methods for Android, however, usually do not consider the use of reflection and cannot detect the malicious behavior hidden by reflection. To further improve software quality, this study proposes a new test case generation method for Android, which uses reflection features to construct a multi-grain model of Android applications. Meanwhile, it analyzes reflection relationships to generate call paths that can reach reflection and employs the adaptive genetic algorithm to generate test cases that cover reflection paths to test Android applications with reflection features. For verification, the effectiveness of this method is evaluated in terms of the effectiveness of the multi-grain model of Android applications and the efficiency of the test method. The experimental results reveal that the automatic test case generation method for Android, which is based on reflection features, is more effective and efficient in detecting reflection.
Abstract: To realize the information-based management of geological archives, this study constructs an information system for geological archives, which is based on Spring Boot microservices architecture and is in systematic combination with the service Gateway and Consul registry. In the research and development process, the development mode of front and rear end separation is employed, and the main part of the front page is developed through Layui. By the Spring Boot framework, a back-end microservice example is built. With both the relational database MySQL and non-relational database Redis as the storage carrier of system data, functional modules are established, such as user management, archive warehousing, archive lending and return, and OCR image recognition. In this system, the geological archives are stored electronically, which promotes the sharing and unified transfer of resources, reduces the maintenance workload of personnel, improves work efficiency, and provides a reference for the data fusion of geological archive information.
Abstract: Text matching is one of the key techniques in natural language understanding, and its task is to determine the similarity of two texts. In recent years, with the development of pre-trained models, text-matching techniques based on pre-trained language models have been widely used. However, these text matching models still face the challenges of poor generalization ability in a particular domain and weak robustness in semantic matching. Therefore, this study proposes an incremental pre-training and adversarial training method for low-frequency words to improve the effect of the text matching model. The incremental pre-training of low-frequency words in the source domain helps the model migrate to the target domain and enhances the generalization ability of the model. Additionally, various adversarial training methods for low-frequency words are tried to improve the model’s adaptability to word-level perturbations and the robustness of the model. The experimental results on the LCQMC dataset and the text-matching dataset in the real estate domain indicate that incremental pre-training, adversarial training, and the combination of the two approaches can significantly improve the text matching results.
Abstract: Lattice quantum chromodynamics (Lattice QCD) is an important theory and method to study the interaction between microscopic particles such as quarks and gluons. By discretizing the spacetime into a four-dimensional structural grid and defining the basic field quantity of QCD on the grid, researchers can use a numerical simulation method to study hadron interactions and properties from the first principle. However, the computation in this process is time-consuming, and large-scale parallel computing is required. The fundamental module of the Lattice QCD computation is the Lattice QCD solver which is the main hot spot of the program running. This work studies the realization and optimization of Lattice QCD solver from a domestic heterogeneous computing platform and proposes a design method of Lattice QCD solver, which realizes BiCGSTAB solver and significantly reduces the iteration numbers. With the odd/even pre-processing technology, the study reduces the computing scale of the problem and optimizes the Dslash module’s memory access in terms of the characteristics of a domestic heterogeneous accelerator. Experimental tests show that the speedup ratio of the solver is about 30 times higher than that of the unoptimized one, which provides a useful reference for the performance optimization of Lattice QCD software of domestic heterogeneous supercomputers.
Abstract: This study aims to detect traffic signs accurately and in real time, reduce traffic accidents, and promote intelligent transportation. An improved YOLOv5s detection algorithm, YOLOv5s-GC, based on computer vision technology is designed to solve the problems of insufficient accuracy, large weight files, and slow detection speed of existing detection models for road traffic signs. Firstly, data is enhanced by copy-paste and then sent to the network for training to improve the detection ability of small targets. Then, Ghost is introduced to build the network, reducing the parameters and calculation amount of the original network, and realizing a lightweight network. Finally, the coordinate attention mechanism is added to the backbone network to enhance the representation and positioning of the attention target and improve detection accuracy. The experimental results show that in comparison with the YOLOv5s, the number of parameters of the YOLOv5s-GC network model is reduced by 12%; the detection speed is increased by 22%; the average accuracy reaches 94.2%. The YOLOv5s-GC model is easy to deploy and can meet the speed and accuracy requirements of traffic sign recognition in actual autonomous driving scenarios.
Abstract: Relevant information of railway accidents, existing in the form of accident overview texts, is of great significance to railway safety work. However, due to the lack of effective information extraction methods, the knowledge of railway accidents scattered in the texts has not been fully utilized. Named entity recognition is an important subtask of information extraction, and there are few studies on named entity recognition of accidents. A named entity recognition model fused with character position features is proposed for the named entity recognition of railway accidents. The model obtains the character position features through a fully connected neural network. It merges them with the character vectors at the semantic level as the final vector representation of the characters, which is then input to the BiLSTM-CRF model to obtain the optimal label sequence. The experimental results show that the accuracy, recall, and F1 value of the model on the named entity recognition of railway accident texts are 93.29%, 94.77%, and 94.02% respectively. This model yields better effects than traditional models and lays a foundation for the construction of a railway accident knowledge graph.
Abstract: Given the poor robustness and low security of traditional watermark-based copyright protection of digital images, a zero watermarking algorithm based on multiple features and chaotic encryption is proposed to improve the differentiability of the zero watermarks of different images. Specifically, global and local perspectives are utilized to extract the five-dimensional features of the image, including its mean feature, variance feature, skewness feature, kurtosis feature, and histogram of oriented gradients (HOG) feature. Then, the watermarked image is encrypted by the proposed block scrambling method based on chaotic mapping. Finally, zero watermarking information is constructed with the multiple features extracted and the scrambled watermarks. The copyright authentication process starts by extracting the multiple features. Then, the encrypted watermarks are obtained according to the zero watermarking information and decrypted, thereby achieving copyright authentication. The experimental results show that the proposed method has high efficiency, high security, and strong anti-attack capability. Integrating various properties of the digital image as the features, the proposed zero watermarking algorithm based on multiple features and chaotic encryption is stable and more robust. In addition, the proposed block scrambling method based on chaotic mapping improves the security of the watermarked images. The proposed algorithm and method effectively solve the problems of the poor robustness and low security of image watermarks.
Abstract: Given the problem that in the field of medical plastic surgery, customers can not directly grasp the postoperative effect of plastic surgery before the surgery, this study proposes a three-dimensional (3D) face reconstruction and editing system for medical plastic surgery. Specifically, the system marks the feature points of the pictures uploaded by the user, aligns the input images with the 3D morphable model (3DMM), and then inputs the processed images into the pre-trained 3D face reconstruction network to obtain the 3D face model corresponding to the input images. After the system loads and renders this model, the user can edit the cheeks, bridge of the nose, and chin of the model to simulate the plastic surgery, and he/she can also save the model to view the diagnosis results. Finally, the reliability of the reconstruction effect, plastic surgery effect, and diagnosis results are tested. The experimental results show that the system is effective in reconstructing young and middle-aged faces, offering reconstructed models highly similar to the input images. The parts of the model remain smooth and natural after the plastic surgery, which means the effect of simulating the plastic surgery is achieved. After the correct face size is determined, the plastic surgery recommendations suggested by the diagnosis results are at the mm level, indicating that the plastic surgery results are highly reliable.
Abstract: An improved SURF-based image matching algorithm is proposed for the traditional SURF image matching algorithm which has the problems of complex computing data, time-consuming, and poor matching accuracy. Firstly, the traditional SURF algorithm is employed to extract the feature points of the image to be matched, and then the 64-dimensional descriptor of SURF is reduced to 20 dimensions by replacing the rectangular area with a circular area. Secondly, the KNN algorithm is utilized to bidirectionally match the feature points of the image to be matched, and the matching pair set of bidirectional initial feature points is obtained. Finally, the mismatching pairs of initial matching points are eliminated bidirectionally by the RANSAC algorithm. The experimental results show that the proposed algorithm reduces the detection time, improves the matching accuracy, and has strong robustness.
Abstract: Map matching is the process of mapping the original global positioning system (GPS) trajectory data of vehicles into the actual road network, and retrieving candidate road sections for GPS trajectory points is the primary link of this process. However, retrieval methods directly affect the accuracy and efficiency of map matching. In this study, a road section retrieval method based on the floating grid is proposed for GPS trajectory data sampled at a low frequency in an urban road network environment. This method resorts to GeoHash grid encoding and floating GeoHash grid to retrieve candidate road sections for trajectory points. Then, to verify the feasibility of the method, this study applies the hidden Markov model, the incremental method, and the Viterbi algorithm to calculate the local optimal solution, with due consideration of the topological structure of the road network and the time-space constraints on the trajectory. Finally, the greedy strategy is employed to obtain the global optimal matching path from the local optimal solution through successive extension.
Abstract: To minimize the performance difference between neural machine translation (NMT) and human translation and solve the problem of insufficient training corpora, this study proposes an improved NMT method based on the generative adversarial network (GAN). First, the sentence sequence of the target end is added with small noise interference, and then the original sentence is restored by the encoder to form a new sequence. Secondly, the results of the encoder are presented to the discriminator and decoder for further processing. In the training process, the discriminator and the bilingual evaluation understudy (BLEU) objective function are employed to evaluate the generated sentences, and the results are fed back to the generator to instruct its learning and optimization. The experimental results demonstrate that compared with the traditional NMT model, the GAN-based model greatly improves the generalization ability and translation accuracy of the model.
Abstract: In the past half-century, with the development of computer technology, neural networks have been widely used in many fields such as images, speeches, and decision-making. To improve the accuracy of neural networks, different scholars have designed a large number of network structures, and thus neural networks have become more and more complex and multi-parametric. As a result, the training process of neural networks has strong non-convexity, and different initial parameters of the same network often train different models. To more accurately describe the performance of two networks, predecessors proposed to evaluate the distribution of the performance of different random seeds on different models trained by the same network through the statistical method of stochastic dominance. On this basis, this study believes that the distribution of the performance of different models on different samples in a test set is also worthy of attention, and thus the stochastic dominance method is applied to compare the distribution of the performance of different models on different samples. Through the experiments on the networks applied in image segmentation, this study finds that for the two models trained by different networks, although one of them has certain advantages in the performance score, it may show stronger dispersion on different samples in the test set. The practical application, however, requires the neural network model with a better performance score and dispersion as small as possible. The stochastic dominance method can effectively compare different models for the selection of a more suitable one.
Abstract: To solve the security problem of Hyperledger Fabric caused by the use of fixed endorsement nodes for endorsement, this study proposes an optimization scheme based on a verifiable delay function for the endorsement policy of Hyperledger Fabric. Considering that the verifiable delay function cannot be calculated in parallel but can be efficiently verified, a Fabric transaction model that anonymously and randomly selects endorsement nodes is designed to enhance the security of Fabric transaction endorsement. The experimental comparison between the optimized scheme with the original one verifies that the optimized scheme not only enhances security but also improves efficiency.
Abstract: Using traditional k-anonymization techniques to achieve privacy protection in social networks is faced with problems such as single clustering criterion and under-utilization of data and information in the graph. To solve this problem, this study proposes an anonymization technique measuring the similarity of the node 1-neighbor graph based on the Kullback-Leibler divergence (SNKL). The original graph node set is divided according to the similarity of node 1-neighbor graph distribution, and the graph is modified according to the divided classes so that the modified graph satisfies k-anonymity. On this basis, the anonymous release of the graph is implemented. The experimental results show that compared with the HIGA method, the SNKL method reduces the amount of change in the clustering coefficients by 17.3% on average. Moreover, the overlap ratio between the importance nodes of the generated anonymous graph and those of the original graph is maintained at more than 95%. In addition to protecting privacy effectively, the proposed method can significantly reduce the changes brought to the structural information in the original graph.
Abstract: Focusing on multi-attribute group decision-making with interval-valued intuitionistic fuzzy numbers as the attribute values, this study uses the grey correlation coefficient method and the information entropy theory to determine the weights of experts and attributes and builds a comprehensive evaluation cloud model through information aggregation. For this purpose, the influences of fuzziness and randomness on the process and results of group decision-making are taken into account, and the cloud model theory and the characteristics of interval-valued intuitionistic fuzzy numbers are leveraged. Different from the traditional ranking method for interval-valued intuitionistic fuzzy numbers, this study uses the 3En rule of the cloud model to convert interval-valued intuitionistic fuzzy numbers into the cloud and determines the comprehensive evaluation value and hesitation degree of the scheme through cloud similarity. Then, the decision-making schemes are comparatively analyzed. The results show that the proposed method can make decisions scientifically and effectively and thereby provide a scientific basis for decision-makers.
Abstract: Considering the problems of low segmentation efficiency of traditional image segmentation methods, complex and diverse features of remote sensing images, and limited segmentation performance in complex scenes, an improved U-Net model is proposed on the basis of the U-Net network architecture, which can satisfactorily extract the features of remote sensing images while maintaining efficiency. First, EfficientNetV2 is used as the encoding network of U-Net to enhance the feature extraction ability and improve the training and inference efficiency. Then, the convolutional structural re-parameterization method is applied in the decoding network and is combined with the channel attention mechanism to improve the network performance without increasing the inference time. Finally, the multi-scale convolution fusion module is employed to improve the feature extraction ability of the network for objects with different scales and the utilization of context information. The experiments reveal that the improved network can not only improve the segmentation performance of remote sensing images but also promote segmentation efficiency.
Abstract: Underwater robots with vision systems cannot operate without the accurate segmentation of underwater objects, but the complex underwater environment and low scene perception and recognition accuracy will seriously affect the performance of object segmentation algorithms. To solve this problem, this study proposes a multi-object segmentation algorithm combining YOLOv5 and FCN-DenseNet, with FCN-DenseNet as the main segmentation framework and YOLOv5 as the object detection framework. In this algorithm, YOLOv5 is employed to detect the locations of objects of each category, and FCN-DenseNet semantic segmentation networks for different categories are input to achieve multi-branch and single-object semantic segmentation. Finally, multi-object semantic segmentation is achieved by the fusion of the segmentation results. In addition, the proposed algorithm is compared with two classical semantic segmentation algorithms, namely, PSPNet and FCN-DenseNet, on the seabed image data set of the Kaggle competition platform. The results demonstrate that compared with PSPNet, the proposed multi-object image semantic segmentation algorithm is improved by 14.9% and 11.6% in MIoU and IoU, respectively. Compared with the results of FCN-DenseNet, MIoU and IoU are improved by 8% and 7.7%, respectively, which means the proposed algorithm is more suitable for underwater image segmentation.
Abstract: Given the problem that no image datasets of defective cloth with defect location information are available for the training of the automatic detection model for cloth defects in industrial production, this study proposes an image generation model EC-PConv with defect location information for defective cloth, and it uses an improved partial convolutional network as its basic framework. This model adopts a feature extraction network for small-scale defects, splices the extracted defect texture features with the blank mask to obtain a mask with position information and defect texture features, and generates an image with defect position information in a repaired way for the defective cloth. Furthermore, a hybrid loss function integrating the mean squared error (MSE) loss is proposed to generate clearer defect textures. The experimental results show that compared with the latest generative adversarial network (GAN) generation model, the proposed model reduces the Frechet inception distance (FID) score by 0.51 and improves the precision P and mean average precision (MAP) values of the generated image of the defective cloth in the cloth defect detection model by 0.118 and 0.106, respectively. This method is more stable than other algorithms in generating images of defective cloth and can generate images of defective cloth that contain defect location information and have higher quality. Therefore, it can effectively solve the problem that no training datasets are available for the automatic detection model for cloth defects.
Abstract: With the rapid development of high-tech with each passing day, the cross fusion and deep correlation among the Internet of Things, big data, and artificial intelligence are implemented. The Internet of Things is fully integrated into all aspects of our life and work as well as social development. At present, the most widely used and mainstream protocol of the Internet of Things is the message queuing telemetry transport (MQTT) protocol, whose inherent advantages of low overhead and low bandwidth have contributed to the access of a large number of Internet of Things devices to the network. However, in the era of the Internet of Everything, “freedom, controllability, safety, and credibility” are the concepts and criteria of industrial development. Many researchers have proposed MQTT-based design schemes for security algorithms. Regarding the paper titled “Data encryption transmission algorithm Based on MQTT”, however, its core algorithm is found to be at risk of key leakage. Therefore, this study points out the defects of this core algorithm and proposes three MQTT-SE algorithms respectively based on symmetric encryption, public key, and mutual verification of public key certificates. These algorithms can achieve the purpose of high-performance and safe encryption transmission even in a low performance MQTT transmission environment.
Abstract: Discipline construction is the core of the development of colleges and universities. With the deepening and strengthening of discipline construction in colleges and universities, the information on discipline construction increases continuously. Nevertheless, the results of discipline construction can not be effectively managed in the manner of discrete document organization, which is not conducive to subsequent analysis and evaluation. To solve this problem, this study focuses on the construction and further application of discipline construction-oriented knowledge graphs. For this purpose, events are extracted from discipline construction texts by the BERT-BiLSTM-CRF model, and related knowledge is supplemented by the crawler. Then, the property graph model is selected to store knowledge, and a preliminary discipline construction-oriented knowledge graph is thereby built. Subsequently, this knowledge graph is availed to build a visualization system for discipline construction, and the minimum Steiner tree algorithm is adopted for the application of intelligent question answering. Finally, the validity of the proposed method is verified by experimental analysis of the methods of discipline construction-oriented event extraction and intelligent question answering.
Abstract: Multi-modal knowledge graph (MMKG) is a new research hotspot in artificial intelligence in recent years. This study provides a construction method for multi-modal domain knowledge graphs to solve the problem that the domain knowledge system of computer science is large and decentralized. Specifically, a systematic MMKG is constructed by crawling the relevant multi-modal data of computer science. However, the construction of an MMKG needs a lot of manpower and material resources. In response, this study trains a model of joint extraction of entities and relations based on the LEBERT model and relation extraction rules and ultimately implements an MMKG of the computer science domain that can automatically extract relation triples.
Abstract: As the detection result lacks interpretability, the Android malware detection is analyzed in terms of interpretability. This study proposes an interpretable Android malware detection method (multilayer perceptron attention method, MLP_At) comprehensively using the multilayer perceptron and attention mechanism. By extracting permissions and application programming interface (API) features from Android malware, it performs data preprocessing on the proposed features to generate feature information, and multilayer perceptrons are utilized for learning features. Finally, the learned data is classified by the BP algorithm. The attention mechanism is introduced in the multilayer perceptron to capture sensitive features and generate descriptions based on the sensitive features to explain the core malicious behavior of the application. The experimental results show that the proposed method can effectively detect malware and the accuracy is improved by 3.65%, 3.70%, and 2.93% compared with that of SVM, RF and XGboost, respectively. The method can accurately reveal the malicious behavior of the software and can also explain the reasons why samples are misclassified.
Abstract: Although deep reinforcement learning can solve many complex control problems, it needs to pay the cost of a large number of interactions with the environment, which is a major challenge for deep reinforcement learning. One of the reasons for this problem is that it is difficult for an agent to extract effective features from a high-dimensional complex input only by relying on the loss of value function. As a result, the agent has an insufficient understanding of the state and cannot correctly assign value to the state. Therefore, this study proposes a regularized predictive representation learning (RPRL) method combining forward state prediction and latent space constraint to make agents know the environment and improve the sample efficiency of reinforcement learning. The method helps agents to learn and extract state features from high-dimensional visual observations to improve the sample efficiency of reinforcement learning. The forward state transfer loss is used as the auxiliary loss so that the features learned by agents contain dynamic information related to environmental transition. At the same time, the state representation of latent space is regularized on the basis of forward prediction, which further helps the agent to learn the smooth and regular representation of the high-dimensional input. In DeepMind control (DMControl) environment, the proposed method achieves better performance than other model-based methods and model-free methods with representation learning.
Abstract: As new deep learning models, graph neural networks are widely used in graph data and promote various applications, such as recommendation systems, social networks, and knowledge graphs. Most existing heterogeneous graph neural models usually predefine multiple metapaths to capture composite relationships in heterogeneous graphs. However, some models usually consider one metapath during the feature aggregation, leading to models only learning neighbor structure but ignoring the global correlation of multiple matapaths. Others omit intermediate nodes and edges along the metapath, which means models cannot learn the semantic information in each metapath. To address those limitations, this study proposes a new model named Metapath-based Graph Transformer Neural Network (MaGTNN). Specifically, MaGTNN first samples heterogeneous graph as metapath-based multi-relation graph and then uses the proposed position encoder and edge encoder to capture the semantic information in a metapath. Subsequently, all the matapath-based neighbor information is aggregated to the target node through their similarity, which is calculated by the improved graph Transformer layer. Extensive experiments on three real-world heterogeneous graph datasets for node classification and node clustering show that MaGTNN achieves more accurate prediction results than state-of-the-art baselines.
Abstract: Different from ordinary object detection tasks, the difficulty of detecting tile surface defects lies in the detection of unconventional size objects, such as small-sized objects and objects with large aspect ratios. To solve these two problems, this study proposes a new type of tile surface defect detection algorithm based on improved Cascade R-CNN. To improve the detection ability for small defects, the model in this study uses the lateral connection structure to fuse the semantic information of the upper and lower layers and applies the dilated convolution with switchable dilation rates to increase the receptive field of the model. To improve the detection ability for defects with large aspect ratios, the proposed model introduces an offset field on the standard convolution to better extract the object feature information. In addition, the model adjusts the size and length of the pre-selected anchor box in the Cascade R-CNN framework. The experimental results show that on the dataset collected from the tile factory, the mean average precision (mAP) of the proposed algorithm reaches 73.5%, which is 9.7% higher than that of the Cascade R-CNN model before improvement. The experimental code of this study is available at: https://github.com/mashibin/Ceramic-tile-defect-detection.
Abstract: Smart city is a new type of smart city transformation under the ternary organic integration of social space, physical space, and information system. It uses a new generation of information technology to optimize the city system, enhance the city’s quality and comprehensive competitiveness, and achieve sustainable development. In recent years, although the construction of smart cities has also been attached increasing importance, it has not been as smooth as the smart transformation of other fields. The construction process is still faced with many problems, which restricts the city’s development to a certain extent. This study takes the urban-scale Happiness Forest Belt building as an example, outlines a smart operation management and control platform developed for this building, and presents a condensed reference technical framework for a smart platform to meet its needs in aspects of digitization, visualization, smartness, and open development frameworks. Finally, some suggestions are proposed on the basis of this experience for constructing an intelligent city operation management and control platform, and its potential critical technical requirements are put forward.
Abstract: The structure of cars determines that there are a large number of blind spots around them, and thus a driver cannot make accurate judgments on the surrounding environment, which is not conducive to safe driving. The holographic transparent image can provide the driver with information on all blind spots around and under the vehicle to assist in safe driving. To solve the problem of obvious seams in image stitching, a distance-based Alpha image stitching algorithm is proposed, and a three-dimensional (3D) model is redesigned for the stitching algorithm. The transparent chassis function is optimized in the following three aspects. The bicycle model algorithm is improved to reduce the computational complexity on the premise of no image effect. The table look-up method is used to improve the accuracy of the conversion of the steering wheel angle into the wheel angle, and the problem of stitching dislocation between the transparent chassis and the surroundings is solved. The method of layered rendering is adopted to optimize the seam problem of the transparent chassis function. The experiments indicate that this technology can effectively improve the rendering effect.
Abstract: Monocular depth estimation is a fundamental problem in computer vision, and the patch-match and plane-regularization network (P2Net) is one of the most advanced unsupervised monocular depth estimation methods. As the nearest neighbor interpolation algorithm, the upsampling method adopted by the depth prediction network of P2Net, has a relatively simple calculation process, the predicted depth maps have a poor generation quality. Therefore, the residual upsampling structure based on multiple upsampling algorithms is constructed in this study to replace the upsampling layer of the original network for more feature information and higher integrity of the object structure. The experimental results on the NYU-Depth V2 dataset reveal that compared with the original network, the improved P2Net based on the transposed convolution, bilinear interpolation, and PixelShuffle can reduce the root mean square error (RMSE) by 2.25%, 2.73%, and 3.05%, respectively. The residual upsampling structure in this study improves the generation quality of the predicted depth maps and reduces the prediction error.
Abstract: This study proposes a deep Q-network (DQN) algorithm based on the K-nearest neighbor (KNN) algorithm (K-DQN) for the energy consumption prediction of buildings. When using the Markov decision process to model the energy consumption of buildings, the K-DQN algorithm shrinks the original action space to improve the prediction accuracy and convergence rate considering large-scale action space problems. Firstly, the original action space is evenly divided into multiple sub-action spaces, and the corresponding state of each sub-action space is regarded as a class to construct the KNN algorithm. Secondly, actions of the same sequence in different classes are denoted by the KNN algorithm to shrink the original action space. Finally, state class probabilities and original states are combined by K-DQN to construct new states and help determine the meaning of each action in the shrunken action space, which can ensure the convergence of the K-DQN algorithm. The experimental results indicate that the proposed K-DQN algorithm can achieve higher prediction accuracy than deep deterministic policy gradient (DDPG) and DQN algorithms and take less network training time.
Abstract: The combination of kernel principal components analysis (KPCA) and control limits (CLS) based on Gaussian distribution will undermine the performance. The fault detection and identification method for nonlinear process based on kernel principal components analysis-kernel density estimation (KPCA-KDE) is proposed. kernel density estimation (KDE) technology is adopted to estimate the CLS based on KPCA for nonlinear process monitoring. According to the detection rate of all 20 faults in KPCA and KPCA-KDE, KDE has a higher fault detection rate than the corresponding method based on Gaussian distribution. In addition, KDE-based detection delay is equal to or lower than other methods. By changing the bandwidth and the number of reserved pivots during the fault detection, KPCA records a larger FAR while the KPCA-KDE does not record any false alarms. The application on the Tennessee Eastman (TE) process shows that KPCA-KDE has better monitoring performance in sensitivity and detection time than KPCA based on Gaussian CLS.
Abstract: The security and efficiency of cloud data storage are urgent issues to be solved in cyberspace security. Therefore, a new ciphertext retrieval model is proposed in the study, and on this basis, the ElGamal homomorphic cipher algorithm and SM4 block cipher algorithm are used to design a cloud ciphertext storage and retrieval solution based on hybrid homomorphic encryption. The retrieval solution can ensure data security during data uploading, retrieving, and downloading and can be applied to personal cloud USB drives and other application scenarios. Moreover, the correctness and safety of the scheme are analyzed and proved through experiments. The experimental results reveal that the scheme can assure correct retrieval results with high efficiency while ensuring data security.
Abstract: To improve the efficiency and accuracy of rail surface defect detection, a rail surface defect detection algorithm based on background difference and maximum entropy is proposed. Firstly, the background model of the rail images is built, and the original images are differentiated from the background images to avoid the influence of illumination change and uneven reflection and accurately highlight the defect area. Then, the improved genetic algorithm is combined with the maximum entropy method to seek the best segmentation threshold and binarize the difference graph. The operational speed of the maximum entropy method is accelerated by the improved genetic algorithm. Finally, the binary images are filtered to complete the segmentation of rail surface defects. The simulations indicate that this method can segment defects quickly and accurately, and the precision, recall, and accuracy are 88.6%, 93.4%, and 90.6%, respectively.
Abstract: In view of a large quantity of parameters in the Inception-v3 network, this study proposes an effective gesture image recognition method, which can meet the needs of high-precision gesture recognition with few model parameters. In this study, the structure of Inception-v3 is used to redesign the Inception module of the original Inception-v3 to reduce the number and difficulty of learning parameters, and with the residual connection, the integrity of information is protected while the network degradation is prevented. The attention mechanism module is introduced to make the model focus on useful information and dilute useless information, and to a certain extent, it also prevents the overfitting of the model. Moreover, the feature fusion is carried out between the up-sampling and the low-level feature in the model, and the fused feature has better discrimination than the original input feature, which further improves the accuracy of the model. The experimental results indicate that the quantity of the parameters in the improved Inception-v3 network is only 1.65 M, and it has higher accuracy and faster convergence speed. Then, the ASL sign language dataset and the Bangladesh sign language dataset are jumbled separately, and the training set and validation set are divided at a ratio of 4:1. The recognition rates of the improved Inception-v3 on the ASL sign language dataset and Bangladesh sign language dataset are 100% and 95.33%, respectively.
Abstract: In recent years, industry-university-research cooperation has become an important factor promoting industrial upgrading and economic development. Industry-university-research services are conducive to quickly integrating multiple resources, improving innovation efficiency, and enhancing the comprehensive competitiveness of enterprises. Governments at all levels have also launched various policies to support industry-university-research cooperation. However, a wide variety of enterprises, as an important part of such cooperation, find it difficult to collect and organize befitting supporting policies efficiently during the application process. Therefore, this study employs the technology of artificial intelligence-based text analysis to design and implement a policy matching system for industry-university-research services. This system can analyze and preprocess various policies from different sources, thereby enabling enterprises to quickly obtain supporting policies that match their specific conditions and ultimately saving manpower for the enterprises and improving their efficiency in applying for a project.
Abstract: In the engineering field, operators need to face complex information interfaces with unevenly distributed stimuli and perform related interactive tasks. Visual attention allocation of operators has been proved to be closely related to task performance. However, the potential connection between visual attention allocation by multi-priority stimuli based on different information allocation strategies and task performance in complex interfaces requires further investigation. In this study, task performance and visual behavior of operators under different load conditions are studied on the basis of the experiment of the multi-priority attention allocation strategy. The experimental results indicate that the differential allocation strategy and information priority division improve the task performance, and the visual behavior differs significantly under different allocation strategies and priorities and is affected by mental loads. This conclusion can provide a reference for the design and optimization of human-computer interfaces and thus improve the task performance of operators.
Abstract: As Bayesian deep learning (BDL) combines the complementary advantages of the Bayesian method and deep learning (DL), it becomes a powerful tool for uncertainty modeling and inference of complex problems. In this study, a BDL framework based on t distribution and the cyclic stochastic gradient Hamiltonian Monte Carlo sampling algorithm is constructed, and a measure of uncertainty is given in view of data uncertainty and model uncertainty. To verify the validity and applicability of the framework, this study constructs corresponding BDL models based on the artificial neural network (ANN), convolutional neural network (CNN), and recurrent neural network (RNN) separately and applies these models to the prediction of 15 global stock indices. The empirical results reveal that 1) the framework is applicable under ANN, CNN, and RNN, and the prediction effect of all indices is excellent; 2) in terms of prediction accuracy and applicability, the BDL models based on t distribution have significant advantages over those based on normal distribution; 3) the MAE under a given uncertainty threshold is better than the original MAE, which indicates that the measure of uncertainty defined in this study is effective and is of great significance to uncertainty modeling. In view of the advantages of the BDL framework in forecasting accuracy, easy to expand and providing measurement of forecasting uncertainty, it has a broad application prospect in finance and other fields with complex data characteristics.
Abstract: Solving expensive optimization problems is often accompanied by computational cost disasters. To reduce the number of real evaluations of the objective function, this study uses the ordinal prediction method in the selection of candidate solutions in evolutionary algorithms. The relative quality of candidate solutions is directly obtained through classification prediction, which avoids the need to establish an accurate surrogate model for the objective function. In addition, a reduction method for the ordered sample set is designed to reduce the redundancy of the ordered sample set and improve the training efficiency of the ordinal prediction model. Next, the ordinal prediction is combined with the genetic algorithm. The simulation experiments of the ordinal prediction-assisted genetic algorithm on the expensive optimization test function show that the ordinal prediction method can effectively reduce the computational cost of solving expensive optimization problems.
Abstract: In light of the structural characteristics of the displacement layer and the basic idea of differential fault, this study proposes a differential fault attack method for the eight-sided fortress (ESF) algorithm. In the 30th round, a 1-bit fault is injected multiple times. Various input and output differential pairs are used to obtain different input sets for the S-box according to the differential characteristics of the S-box. Taking the intersection of the sets is a quick way to determine the only possible inputs for the S-box. The round key of the last round can then be obtained through analysis. Similarly, a 1-bit fault is injected in the 29th and 28th rounds many times. With the round key of the last round, the differential characteristics of the S-box are leveraged again to obtain the round keys of the last but one and last but two rounds. About 10 fault ciphertexts are required. After the round keys of three rounds are recovered, the computational complexity of recovering the master key is reduced to 222.
Abstract: Traditional token-based clone detection methods utilize the serialization characteristics of code strings to quickly detect clones in large code repositories. However, compared with the methods based on the abstract syntax tree (AST) and program dependency graph (PDG), traditional methods can hardly detect code clones with large text differences due to the lack of syntax and semantic information. Therefore, this study proposes a token-based clone detection method with semantic information. First, AST is analyzed, and the semantic information of tokens located at the leaf nodes is abstracted using the AST path. Then, a low-cost index is established on the tokens for function names and type roles to quickly filter valid candidate clone fragments. Finally, the similarity between code blocks is judged using the tokens with semantic information. The experimental results on the public large-scale dataset BigCloneBench reveal that this method significantly outperforms the mainstream methods, including NiCad, Deckard, and CCAligner in Moderately Type-3 and Weakly Type-3/Type-4 clones with low text similarity while requiring less detection time on large code repositories.
Abstract: Action recognition aims to make computers understand human actions by the processing and analysis of video data. As different modality data have different strengths in the main features such as appearance, gesture, geometric shapes, illumination, and viewpoints, action recognition based on the multi-modality fusion of these features can achieve better performance than the recognition based on single modality data. In this study, a comprehensive survey of multi-modality fusion methods for action recognition is given, and their characteristics and performance improvements are compared. These methods are divided into the late fusion methods and the early fusion methods, where the former includes prediction score fusion, attention mechanisms, and knowledge distillation, and the latter includes feature map fusion, convolution, fusion architecture search, and attention mechanisms. Upon the above analysis and comparison, the future research directions are discussed.
Abstract: Considering strong noise interference and difficult shadow detection in high-resolution remote sensing images of high-rise buildings, this study proposes a shadow detection method for remote sensing images of high-rise buildings, which is based on the combination of improved threshold segmentation and residual attention networks. Firstly, a threshold segmentation model is built by the improved maximum inter-class and minimum intra-class threshold segmentation algorithm, and on the basis of the connected domain characteristics and end-point positional constraint relationships between contours, the Euclidean metric algorithm is used to repair the broken contours for the shadow contours. Then, the generative adversarial network (GAN) model is used to expand the misjudgment data set. Finally, the residual network is improved, and the attention mechanism is added to the feature map for global feature fusion. In different scenes, the proposed method is compared with the radiation model, histogram threshold segmentation, color model-based shadow detection method, support vector machine (SVM), visual geometry group (VGG) network, Inception, and classification network of residual networks, and the proposed method has a comprehensive misjudgment rate and missed detection rate of 2.1% and 1.5%, respectively. The results reveal that the proposed algorithm can better complete the segmentation and detection of shadow areas, which is conducive to saving human and material resources and assisting staff with their work such as interpreting remote sensing information and establishing remote sensing archives. The proposed method has practical value.
Abstract: In high-performance computing, the huge communication overhead has become one of the main bottlenecks in the improvement of its computing power, and the optimization of communication performance has always been an important challenge. For the communication optimization task, this study proposes a method based on in-network computing technology to reduce the communication overhead. In the Ethernet-based supercomputing environment, this method utilizes the RoCEv2 protocol, programmable switches, and OpenMPI to offload reduction computation to programmable switches, and it supports the two communication modes of Node and Socket. The collective communication benchmark test and the OpenFOAM application test are carried out in a real supercomputing environment. The experimental results indicate that when the number of server nodes reaches a certain scale, compared with the traditional host communication, this method shows better performance improvement in both Node and Socket modes, with the performance in the collective communication benchmark test improved by about 10%–30% and the overall application performance in the application-level test improved by about 1%–5%.
Abstract: Speech emotion recognition (SER) plays an extremely important role in the process of human-computer interaction (HCI), which has attracted much attention in recent years. At present, most SER approaches are mainly trained and tested on a single emotion corpus. In practical applications, however, the training set and testing set may come from different emotion corpora. Due to the huge difference in the distribution of different emotion corpora, the cross-corpus recognition performance achieved by most SER methods is unsatisfactory. To address this issue, many researchers have started focusing on the studies of cross-corpus SER methods in recent years. This study systematically reviews the research status and progress of cross-corpus SER methods in recent years. In particular, the application of the newly developed deep learning techniques on cross-corpus SER tasks is analyzed and summarized. Firstly, the emotion corpora commonly used in SER are introduced. Then, on the basis of deep learning techniques, the research progress of existing cross-corpus SER methods based on hand-designed features and deep features is summarized and compared from the perspectives of supervised, unsupervised, and semi-supervised learning. Finally, the challenges and opportunities in the field of cross-corpus SER are discussed and predicted.
Abstract: Different from the laboratory environment, the scenes of facial expression images in real life are complex, and local occlusion, the most common problem, will cause a significant change in the facial appearance. As a result, the global feature extracted by a model contains redundant information unrelated to emotions, which reduces the discrimination of the model. Considering this problem, a facial expression recognition method combining contrastive learning and the channel-spatial attention mechanism is proposed in this study, which learns local salient emotion features and pays attention to the relationship between local features and global features. Firstly, contrastive learning is introduced. A new positive and negative sample selection strategy is designed through a specific data augmentation method, and a large amount of easily accessible unlabeled emotion data is pre-trained to learn the representation with occlusion-aware ability. Then, the representation is transferred to the downstream facial expression recognition task to improve recognition performance. In the downstream task, the expression analysis of each face image is transformed into the emotion detection of multiple local regions. The fine-grained attention maps of different local regions of a face are learned using the channel-spatial attention mechanism, and the weighted features are fused to weaken the noise effect caused by the occlusion content. Finally, the constraint loss for joint training is proposed to optimize the final fusion feature for classification. The experimental results indicate that the proposed method achieves comparable results to existing state-of-the-art methods on both public non-occluded facial expression datasets (RAF-DB and FER2013) and synthetic occluded facial expression datasets.
Abstract: Docker image is the operating basis of Docker containers. As robust methods of image security detection remain to be developed, containers are subject to various security threats, such as container escape and denial of service attacks, during their operation. To avoid the use of toxic images, this study proposes a detection model for trusted Docker image sources, namely detect trusted Docker image source (DTDIS). In this model, the virtual trusted cryptography module (vTCM) is used to build an image benchmark database and thereby detect whether the local image file has been tampered with. The parent image vulnerability database is utilized to extend the Clair image scanner and thus avoid repeated scanning. File measurement information and vulnerability scanning information are availed to determine whether the Docker image source is credible. Experiments in a cloud environment prove that the proposed model can effectively evaluate the security of Docker images and ensure that users use trusted images.
Abstract: Whether the lung is infected by COVID-19 can be effectively detected from lung computed tomography (CT) images by the computer-aided diagnosis system whose training is based on deep learning. However, the main problem is the lack of high-quality labeled CT images available for training. This study proposes a method of augmenting lung CT images with the generative adversarial network (GAN). Specifically, labels of different infected areas are generated, and Poisson fusion is performed to enhance the diversity of the generated images. Then, image transformation and generation are implemented by training the GAN model to fulfill the purpose of augmenting the CT image. Segmentation experiments based on the augmented data are also carried out to verify the effectiveness of the data generated. The results of the image generation and segmentation experiments both show that the proposed image generation method achieves favorable effects.
Abstract: The vehicle routing problem (VRP) exists extensively in the modern logistics industry, which is an NP-hard problem in combinatorial optimization. Affected by factors such as diverse customer demand and road traffic restrictions, VRP becomes more complex, and it can hardly be solved by the traditional combinatorial optimization methods and operations research methods. In this study, a common VRP with time windows is studied. The waiting time of vehicles is reduced by the adjustment to the priority of customers according to the parameters of time windows. On this basis, several common heuristic algorithms are improved, and 56 common VRPs are tested. The experimental results reveal that the improved savings algorithm can produce good results for capacitated VRPs, and the improved insertion method has superior performance in VRPs with time windows. In addition, the improved heuristic algorithms can make the total distance better than the known optimal value when using more vehicles on the four test cases.
Abstract: In the research of bird sound recognition, the selection of sound features has a great impact on the accuracy of recognition and classification. To improve the accuracy of bird sound recognition, this study starts with the problem that the traditional Mel frequency cepstral coefficient (MFCC) characterizes the high-frequency information in bird sound insufficiently. Feature fusion of MFCC based on Fisher criterion and inverted MFCC (IMFCC) is proposed to obtain a new feature parameter MFCC-IMFCC that can be applied to bird sound recognition to improve the characterization of the high-frequency information in bird sound. Meanwhile, the penalty factor C and the kernel parameter g in the support vector machine (SVM) are optimized by a genetic algorithm (GA), and a GA-SVM classification model is trained. Experiments show that under the same conditions, the recognition rate of the MFCC-IMFCC approach is higher than that of the MFCC one.
Abstract: Traditional terminology standardization schemes based on template matching, artificially constructed features, semantic matching, etc., are often faced with problems such as low terminology mapping accuracy and difficult alignment. Given the colloquial and diverse expression of terminology in medical texts, modules of multi-strategy recall and implication semantic score ranking are used to improve the effect of medical terminology standardization. In the multi-strategy recall module, the recall method based on the Jaccard correlation coefficient, term frequency-inverse document frequency (TF-IDF), and historical recalls is employed. In the implication semantic scoring module, RoBERTa-wwm-ext is adopted as the scoring semantic model. The usability of the proposed method is validated for the first time on a Chinese dataset that is based on the systematized nomenclature of medicine-clinical terms (SNOMED CT) standard and annotated by medical professionals. Experiments show that in the processing of medical knowledge features, the proposed method can achieve favorable results in practical applications of medical terminology standardization and has high generalization and practical value.
Abstract: With the rapid development of the Internet of Things (IoT), the number of IoT devices has grown exponentially, which is accompanied by the increasing attention to IoT security. Generally, IoT devices adopt software attestation to verify the integrity of the software environment, so that system integrity tampering caused by the execution of malicious software can be detected timely. However, the existing software attestation suffers from poor performance in the synchronous attestation of massive IoT devices and the difficulty in extending the general IoT communication protocol. To address these problems, this study proposes a lightweight asynchronous integrity monitoring scheme. The scheme extends the security authentication message of software attestation on the general message queuing telemetry transport (MQTT) protocol and asynchronously pushes the integrity information of devices. It improves not only the security of IoT systems but also the efficiency of integrity attestation and verification. The following three security functions are realized: device integrity measurement in a kernel module; lightweight authentication extension of device identity and integrity based on MQTT; asynchronous integrity monitoring based on MQTT extension protocol. This scheme can resist common software attestation attacks and MQTT protocol attacks and has the characteristics of lightweight asynchronous software attestation and general MQTT security extension. The experimental results of the prototype system of IoT authentication based on MQTT show the high performance of the integrity measurement of IoT nodes, MQTT protocol connection authentication and PUBLISH message authentication, which can meet the application requirements of integrity monitoring of massive IoT devices.
Abstract: Accurate prediction of sea surface temperature (SST) is vital for marine fishery production and the prediction of marine dynamic environment information. The traditional numerical prediction methods have high calculation costs and low time efficiency. However, the existing data-driven SST prediction methods mainly target the single observation point and fail when it comes to a sea region composed of multiple observation points. The existing regional SST prediction methods still have a long way to go in prediction accuracy. Therefore, we propose a regional SST prediction method based on XGBoost and PredRNN++ (XGBoost-PredRNN++). The method firstly converts SST data into gray images and then extracts the time characteristics of each point by the XGBoost model. On this basis, the CNN network is utilized for fusing the time characteristics into the original SST data, and the spatial dependence is extracted at the same time. Finally, the latest time series prediction model PredRNN++ is adopted to extract the temporal and spatial correlations among SST data to achieve the high-precision prediction of regional SST. The experimental results show that the high prediction accuracy and efficiency of the proposed method are superior to those of the existing methods.
Abstract: In large industrial plants, due to a wide variety and a large number of equipment control switches, the complexity of operating procedures and the subjectivity of human judgment may lead to operational errors and cause serious consequences in the daily operation and maintenance process. To assist operators in accurately judging whether the state of an equipment switch is correct, an improved Faster R-CNN algorithm is proposed for state recognition of equipment switches. Firstly, the dilated residual network (ResNet) is used as the feature extraction network, and the multi-branch dilated convolution is introduced into ResNet50 to fuse the information of different receptive fields. Secondly, the feature pyramid network is improved by the addition of a bottom-up feature enhancement branch to the original network, which is used to integrate multi-scale feature information. Then, the K-means++ algorithm is applied to cluster bounding boxes of switches, and the size of proposals for equipment switches is designed. Finally, the non-maximum suppression (NMS) algorithm is replaced with Soft-NMS to reduce the influence of switch overlap on the detection effect and enhance the performance of suppressing the overlapping proposals. On a switch state dataset, the mean average precision (mAP) of the improved Faster R-CNN reaches 91.5%. Moreover, it has been applied to assist state recognition of equipment switches in the daily operation and maintenance of pumped-storage power stations to meet the needs of intelligent supervision in complex scenarios.
Abstract: With the continuous development of science and technology, medical diagnosis technology also makes continuous progress. Ultrasound technology, as a means of medical diagnosis, has been widely used in various medical fields. It has been widely recognized by doctors and patients because it is harmless to the human body and can dynamically and clearly show the health state of human tissues and organs. With the continuous development of ultrasound technology, people have higher requirements for the quality of ultrasound imaging. Due to the limitations of the materials of ultrasonic probes, such as for the manufacturing of ceramic transducer, and the compromise scheme of low-channel scanning adopted to reduce the cost and frame rate, the caused noises and artifacts will block the useful information of human tissues and organs, which will seriously affect doctors’ auxiliary diagnosis. In the field of ultrasound, how to enhance images and videos and suppress artifacts has become an important challenge. This study describes several filtering algorithms for artifact suppression in the spatial domain and their limitations and proposes an artifact suppression algorithm based on the frequency domain, which can well suppress the periodic artifact in real-time ultrasonic imaging. Firstly, this study simulates the periodic artifact with a sine wave to highlight its characteristics in the frequency domain. Then, the ultrasonic image is subjected to a two-dimensional Fourier transform into the frequency domain to suppress these artifacts. Because these artifacts are periodic, they have obvious characteristics in the frequency domain. The set corresponding to these artifacts in the frequency domain is found through the algorithm model of sliding window scanning combined with a threshold. Next, according to the dynamic range of the frequency domain and the given threshold, the points of these suspected artifacts in the set are depressed. Finally, the ultrasonic image is transformed into the spatial domain by inverse Fourier transform to obtain the processed image. This method can improve the suppression of periodic artifacts in ultrasonic images and retain useful information, thus able to enhance the accuracy of doctors’ judgment regarding human organ conditions.
Abstract: Skeleton data is compact and robust to environmental conditions for hand gesture recognition. Recent studies of skeleton-based hand gesture recognition often use deep neural networks to extract spatial and temporal information. However, these methods are likely to have problems such as complicated computation and a large number of model parameters. To solve this problem, this study presents a lightweight and efficient hand gesture recognition model. It uses two spatial geometric features calculated from skeleton sequences and automatically learned motion trajectory features to achieve hand gesture classification with convolutional networks alone as its backbone network. The proposed model has a minimum number of parameters as small as 0.16M and a maximum computational complexity of 0.03 GFLOPs. This method is also evaluated on two public datasets, where it outperforms the other methods that use skeleton modality as input.
Abstract: The current operation mode of tower cranes has the problems such as high safety risks and low utilization of operators on sites. To solve the problems, a remote control system of tower cranes based on 5G MEC is proposed, which can freely access multiple tower cranes and multiple clients in different geographical positions and perform comprehensive management and control. It ensures that the low time delay based on 5G communications can be realized at the application level by modular design and targeted strategies for forwarding control of status data, control data, and media stream data, which provides a reference for distributed remote control of multiple devices and clients.
Abstract: In recent years, compute-intensive and time delay-sensitive applications such as AR/VR, online games, and 4K/8K ultra-high-resolution videos have been emerging. Due to the limitations of their hardware conditions, some mobile devices are unable to calculate such applications under the time-delay requirements, and running such applications will consume huge energy and reduce the endurance of mobile devices. To solve this problem, this study proposes an edge computing offloading and resource allocation scheme in a Wi-Fi network with the coordination of multiple access points (APs). Firstly, the genetic algorithm is utilized to determine the task offloading decision of users. Then, the Hungarian algorithm is used to allocate communication resources to users with task offloading. Finally, according to the time-delay limit of task processing, the computing resources of mobile edge computing (MEC) servers are allocated to the users with task offloading. The simulations reveal that the proposed task offloading and resource allocation scheme can effectively reduce the energy consumption of mobile devices on the premise of meeting the time-delay limit of task processing.
Abstract: Considering the low signal-to-noise ratio (SNR) and image detail loss caused by additive white Gaussian noise (AWGN), an image denoising model based on the convolutional neural network (CNN) with residual dense blocks is proposed on the basis of the existing CNN algorithms. By introducing a multi-stage residual network and dense connections and using the leaky relu activation function on the whole network, the model can better retain the effective information of images and effectively avoid feature loss while removing the noise of different levels of intensity. Compared with the residual learning model of the denoising CNN (DnCNN), the proposed model has an improved peak SNR by about 0.12 dB on the Set12 and Bsd68 test sets and improved structural similarity by about 0.008 6 on average. The test results reveal that the proposed model can fully extract image features, retain image details, and reduce the computational complexity of the network.
Abstract: Crowd behavior recognition has important application value in public safety and other fields. Existing studies have considered the influence of such factors on crowd behavior as crowd emotions, crowd types, crowd density, and social and cultural backgrounds of crowds separately, but few models comprehensively consider these factors, which limits model performance. This study comprehensively considers the correlation between the physical features, social features, emotional and personality features, and cultural background features of the crowd and the influence of the combination of these factors on crowd behavior. As a result, a crowd behavior recognition model that integrates multiple features and time series is proposed. The model uses two parallel network layers to deal with the influence of multi-feature correlation and time-series dependence on crowd behavior separately. Meanwhile, the network layer fuses the structural causal model (SCM) and the causal graph network (CGN) based on the graph neural network (GNN) to improve the interpretability of the model. The experiments on the motion and emotion dataset (MED) and the comparison with other state-of-the-art models demonstrate that the proposed method can successfully identify crowd behavior and outperform the state-of-the-art methods.
Abstract: Traditional fire warning methods have low detection accuracy and cannot give early warnings in time when there is no fire. Therefore, this study proposes an early fire warning algorithm based on deep learning. Firstly, an infrared thermal imager is used to collect infrared images in a specific scene for dataset construction. Secondly, the improved YOLOv4 algorithm is applied for training, and the network weights are obtained. The convolutional attention module is introduced after the three output feature layers of the backbone network to improve the ability of the network to extract key information. Convolutional layers are added to the backbone network and path aggregation network to promote feature extraction capability. Finally, the proposed intelligent fire detection (IFD) algorithm is employed to process the predicted image and evaluate the fire hazard according to the score. The experimental results reveal that the mAP of the improved YOLOv4 algorithm on the dataset reaches 98.31%, which is 2.7% higher than that of the original YOLOv4 algorithm, and the FPS is 37.1 f/s; the accuracy of the IFD algorithm is 93%, and its false detection rate is 3.2%. The proposed early fire warning algorithm has the advantages of high detection accuracy and timely warnings when there is no fire.
Abstract: Although the SemBERT model is an improved version of the BERT model, it has two obvious defects. One is its poor ability to obtain vector representation. The other is that it directly uses conventional features to classify tasks without considering the category of the features. A new feature reorganization network is proposed to address these two defects. This model adds a self-attention mechanism into the SemBERT model and obtains better vector representation with an external feature reorganization mechanism. Feature weights are also reassigned. Experimental data show that the F1 score of the new method on the Microsoft Research Paraphrase Corpus (MRPC) dataset is one percentage point higher than that of the classical SemBERT model. The proposed model has significantly improved performance on small datasets, and it outperforms most of the current outstanding models.
Abstract: With the development of 3D digital virtual humans, speech-driven 3D facial animation technology has become one of the important research hotspots in virtual human interaction. The key parts of the speech-driven 3D facial animation technology include the construction of a speech-visual mapping model and the synthesis of 3D facial animation. Specifically, the characteristics of phoneme-viseme matching methods and speech-visual parameter mapping methods are described. Next, the current methods of building 3D facial models are expounded, and the advantages and disadvantages of different motion control methods are analyzed according to the different representation methods of 3D facial models. Then, the subjective and objective evaluation methods for speech-driven 3D facial animation are expounded. Finally, the future development directions of speech-driven 3D facial animation technology are summarized.
Abstract: This study mainly analyzes the sentiment of user reviews on hotels by investigating the attitudes of users toward hotel configuration and service to help hotels improve the quality of personalized service. Specifically, a pretraining model based on the BiLSTM neural network is built and compared with traditional machine learning algorithms. The experimental results reveal that the analysis accuracy of support vector machines (SVMs) is more stable compared with that of naive Bayes, while the prediction accuracy using the pretraining model is slightly improved compared with that of the previous two. Moreover, an extended dictionary of sentiment, with the basic dictionary as the main part, is constructed for reviews on hotels, and the weights of negatives are weakened to reduce the impact on the classification of sentences with opposite meanings. The basic dictionary and the extended dictionary are used to classify the sentiment of the same corpus obtained, and the comparison of the results indicates that with the extended dictionary, the accuracy of the positive classification and negative classification is 86% and 84%, respectively. This indicates that the classification effect of the extended dictionary is better than that of the basic dictionary.
Abstract: OpenCL is an open source and free heterogeneous computing framework, which is widely used in architecture processors. HXDSP is a domestic DSP chip independently developed by the 38th Research Institute of China Electronic Technology Corporation. To solve the scheduling difficulties and insufficient hardware utilization of the HXDSP heterogeneous computing platform, this work studies the task scheduling system of OpenCL during operation. The automatic task graph extraction method during the operation of OpenCL is designed, and the classic static scheduling algorithm HEFT is improved by the combination of the hardware characteristics of HXDSP and the execution model characteristics of OpenCL. Thus, a heterogeneous dual-granularity earliest finish time (HDGEFT) scheduling algorithm is proposed, and experiments are designed on the HXDSP heterogeneous computing platform for verification. The experimental results reveal that the specially designed scheduling algorithm has great advantages in execution efficiency.
Abstract: For lower complexity of the load sequence, the empirical mode decomposition (EMD) method is used to obtain different components. For shorter training time and a smaller cumulative error caused by component forecasting one by one, the components are reconstructed into high-frequency and low-frequency ones according to the zero-crossing rate of the components. The high-frequency components of the load are forecasted by the temporal convolutional network (TCN) model, whereas the low-frequency ones are forecasted by the extreme learning machine (ELM). The proposed EMD-TCN-ELM model is compared with three individual models TCN, ELM, and long short-term memory (LSTM) and three mixed models EMD-TCN, EMD-ELM, and EMD-LSTM through experiments, and its mean absolute percentage error (MAPE) is reduced by 0.538%, 1.866%, 1.191%, 0.026%, 1.559%, and 0.323%, respectively. The forecasting accuracy of the proposed model is also the highest. Additionally, the proposed model has the shortest training time among the top three models in forecasting accuracy. The above results verify the superiority of the proposed model in load forecasting accuracy and training time.
Abstract: A billet is dispatched from the inventory to the bench by a crane and then from the bench to the front of the furnace through a track. In the past, the billet was pushed onto the track by the manual control of machinery. The automation of this process requires knowledge of the real-time position distribution of billets on the bench for automatic control of the pusher. In this study, the real-time positioning of billets on the bench is achieved by the machine vision method. Specifically, with the U-Net as the basic network, the residual blocks in classic ResNet are used to achieve the accurate segmentation of transverse positions of billets. The experimental results and field application tests indicate that the segmentation accuracy of this method can meet the control requirements of industrial fields.
Abstract: To solve the problems of missing feature extraction by convolutional neural network and insufficient multi-feature extraction of a gesture, this study proposes a static gesture recognition method based on a residual double attention module and a cross-level feature fusion module. The designed residual double attention module can enhance the low-level features extracted by a ResNet50 network, effectively learn the key information, update the weight, and improve the attention to high-level features. Then, the cross-level feature fusion module fuses the high-level and low-level features in different stages to enrich the semantic and location information between different levels in the high-level feature map. Finally, the Softmax classifier of the fully connected layer is used to classify and recognize the gesture image. The experiment is carried out on the American sign language (ASL) dataset. The average recognition accuracy is 99.68%, which is 2.52% higher than that of the basic ResNet50 network. The results show that the proposed method can fully extract and reuse gesture features and effectively improve the recognition accuracy of gesture images.
Abstract: The research on the recognition of abnormal human behavior in video surveillance systems is of great significance. As traditional algorithms are easily affected by the environment and have poor timeliness and accuracy, an abnormal behavior recognition algorithm based on skeleton sequence extraction is proposed. Firstly, the improved YOLOv3 network is used to detect targets and is combined with the RT-MDNet algorithm to track them for target trajectories. Then, the OpenPose model is employed to extract the skeleton sequence of targets in the trajectories. Finally, the spatiotemporal graph convolutional network combined with clustering is applied to recognize the abnormal behavior of the targets. The experimental results indicate that the proposed algorithm has a processing speed of 18.25 fps and recognition accuracy of 94% under a complex background of light changes, which can accurately identify the abnormal behavior of various targets in real time.
Abstract: Probabilistic matrix factorization model, making personalized item recommendations according to a user’s historical interaction information, is one of the classic methods in collaborative filtering. Under the assumption of the traditional matrix factorization model, the similarities among different users cannot be used, and prediction is often inaccurate when outliers occur. A clustering-based probabilistic matrix factorization model with category-related conjugate prior distribution is built with user clustering information. Its parameters are regularized by changing the form of the conjugate prior distribution. Through variational inference, the explicit expressions of variational parameters are theoretically derived, and corresponding rating prediction algorithms are thereby established. Both simulation and real datasets show that the prediction performance of the proposed model is better than that of the benchmark model, and it can provide realistic explanations for users’ rating behavior.
Abstract: This study is conducted to trace the source of sudden river pollution. Specifically, the coupling relationship between forward and reverse mass probability density is used to realize the decoupling of the location, discharge time, and discharge intensity of pollution sources; then, given the one-dimensional water body diffusion model and the monitoring data from the tracer experiment in Truckee River of the United States, a method for tracing the source of sudden river pollution is established on the basis of the improved firefly algorithm (FA). In the method, the monitoring data are divided into a training set and an experimental set, and by the training set data, the improved FA is employed to adjust the hydrological parameters of the river. Then, the adjusted hydrological parameters are used in the experimental set, and data from different monitoring sections are used independently for solutions. Finally, the results are analyzed by variance to exclude the data with large source-tracing errors. The results reveal that the source-tracing results have high accuracy and the ability to correct the monitoring data, which is of certain guiding significance for the actual sudden river pollution.
Abstract: Electroencephalography (EEG) has dynamic, nonlinear and numerically highly random signals. To break the limitations of traditional artificial neural network models in feature extraction and recognition accuracy during EEG recognition, this study proposes a new recognition method, which is based on the KIV model to recognize EEG signals. First, the dynamic characteristics of the KIV model under different stimuli are analyzed through simulation experiments. Then, the KIV model is used to recognize epileptic EEG signals and emotional EEG signals. Without feature extraction during the experiment, multi-channel raw EEG signals are directly input into the KIV model for recognition. The recognition accuracy is about 99.50% and 90.83% on BONN and GAMEEMO datasets, respectively. The results show that the KIV model outperforms existing models in the ability to recognize EEG signals and can provide help for EEG recognition.
Abstract: The traditional generative model ignores the important clues provided by key words in the process of abstract generation, which leads to the loss of key word information, and the generated abstract cannot agree with the original text well. In this study, an abstract generation method is proposed, which takes the pointer-generator network as the framework and integrates BERT pretraining model and key word information.?Firstly, the TextRank algorithm and the sequence model based on the attention mechanism are used to extract key words from the original text, and thus the generated key words can contain more information about the original text.?Secondly, the key word attention is added to the attention mechanism of the pointer-generator network to guide the generation of an abstract.?In addition, we use the double-pointer copy mechanism to replace the copy mechanism of the pointer-generator network and thus improve the coverage of the copy mechanism. The results on LCSTS data sets reveal that the designed model can contain more key information and improve the accuracy and readability of generated abstracts.
Abstract: Multi-attribute data privacy publication fails to balance the difference in attribute sensitivity and computational efficiency. For this reason, HMPrivBayes, a heterogeneous multi-attribute data publishing method with differential privacy based on attribute segmentation, is proposed. Firstly, the spectral clustering algorithm satisfying differential privacy is designed to segment the original data set, in which the similarity matrix is generated by the attribute maximum information coefficient. Secondly, with the help of attribute information, this method uses an improved Bayesian network construction algorithm to build Bayesian networks for each data subset. Finally, HMPrivBayes adds heterogeneous noise disturbance to the attribute joint distribution extracted from the Bayesian network to realize the protection of heterogeneous multi-attribute data, in which privacy budget is allocated based on the normalized risk entropy of attribute. The experimental results show that HMPrivBayes not only reduces the added noise but also improves the computational efficiency of synthetic data.
Abstract: Load forecasting methods emerge one after another to maintain the stability of power grids. However, due to the characteristic difference in the generalization ability of algorithms and model complexity, the applicability of these methods to load forecasting differs. This paper discusses and summarizes the research status of short-term power load forecasting both at home and abroad in the past five years from multiple dimensions, such as experimental data sets, data preprocessing, forecasting algorithms, optimization models, and evaluation methods. Meanwhile, we also present a summary of the advantages, disadvantages, and applicability of various forecasting algorithms, and the development trend of the short-term power load forecasting system is expounded and predicted. This study is expected to provide a reference for the forecasting model selection of power system loads in the future.
Abstract: For the cold start, sparse user feedback, and poor accuracy of similarity measurement in traditional article recommendation methods, this paper proposes contextualized topic Bert (ctBert), an article similarity calculation method that combines Bert with the topic model. The algorithm calculates the similarity scores between the given query and the related articles. The preprocessed articles are input into separate sub-modules for feature extraction and similarity score calculation. The similarity score is combined with the personalization score of the support set to obtain the final score. The algorithm is further improved by integrating single-sample learning into the recommendation framework. The experimental results from three different datasets show that the proposed method improves the NDCG criteria on all three datasets. For example, the NDCG@3 and NDCG@5 criteria improve by 6.1% and 7.2% respectively compared with other methods on the Aminer dataset, which validates the effectiveness of the method.
Abstract: The mainstream all-mappers of next-generation sequencing mostly use the seed-and-extend method. Due to high storage costs or long retrieval time of the long-seed index, most of these algorithms use short seeds, which results in redundant candidate positions and increases the time cost of alignment. We, therefore, propose an all-mapper based on long seeds, and a long-seed hash index with low storage costs and moderate retrieval time is designed. The long-seed hash index limits the hash space through modular operation and uses the Bloom filter to identify different seeds at the same storage location. Long seeds significantly reduce the number of candidate locations and thus lower the time cost in the verification phase. The experiments on human gene sequencing datasets reveal that the proposed all-mapper has higher time efficiency than the existing mainstream all-mappers while maintaining the same accuracy.
Abstract: This paper proposes a new flower pollination algorithm by incorporating the improved teaching-learning-based optimization strategy and dynamic Gaussian mutation to enhance the optimization performance. The algorithm first speeds the convergence through the promotion effect between the optimal individual and other individuals obtained by the improved teaching factor in the teaching mechanism. At the same time, the mutual learning mechanism between individuals is adopted to maintain the diversity of the population, thereby improving the optimization accuracy. Then, when it is detected that the algorithm falls into prematurity, the dynamic Gaussian mutation is carried out on the middle individuals of the population to increase the differences between individuals. In this way, it avoids the prematurity of the algorithm and then improves the comprehensive optimization ability. The optimization results of 16 standard functions are checked by the nonparametric statistical test to prove the effectiveness of the algorithm. Compared with other improved pollination algorithms, this algorithm has significant advantages. Finally, the new algorithm is applied to solve the application problems of telescopic rope, and good optimization results are achieved.
Abstract: Retinal vessel segmentation is vital for assisting doctors in diagnosing ophthalmic diseases, including diabetic retinopathy, macular atrophy, and glaucoma. The attention mechanism is widely used in U-Net and its variants to improve the vessel segmentation performance. For more accurate retinal vessel segmentation and exploration of high-order and global context information, we propose a multi-scale high-order attention network (MHA-Net). The multi-scale high-order attention (MHA) module first extracts multi-scale and global features from the high-level feature maps to compute the initial attention map, enabling the model to handle medical image segmentation with variable scales. Then the high-order attention constructs the attention map through graph transduction followed by the extraction of high-level features at high order. We further embed the proposed MHA module into an efficient encoder-decoder structure for retinal vessel segmentation. Comprehensive experiments are conducted on the color fundus image dataset DRIVE, which indicates that the proposed method improves the accuracy of retinal vessel segmentation effectively.
Abstract: The cutting-edge technology in deep learning is applied to surface defect detection of strip steel for the accuracy improvement in surface defect detection of industrial hot-rolled strip steel. Therefore, a surface defect detection algorithm for hot-rolled strip steel is proposed, which takes Swin Transformer as the backbone feature extraction network and cascaded multi-threshold structure as the output layer. Compared with the deep learning target detection algorithm based solely on convolutional networks, the detection algorithm using the Transformer structure can achieve more accurate detection results. Specifically, first, Swin Transformer is used as the backbone feature extraction network to replace the conventional residual network structure and thus enhance the ability of the feature network to capture the deep semantic information implicit in an image. Secondly, a multi-cascade detection structure is designed, and step-by-step IoU thresholds are set to achieve the balance between detection accuracy and threshold improvement. Finally, training strategies such as soft non-maximum suppression (Soft-NMS), FP16 mixed precision training, and SGD optimizers are employed to accelerate model convergence and improve model performance. The experimental results reveal that the proposed algorithm has better detection performance on the industrial hot-rolled strip steel data set (NEU-DET) than the deep learning algorithms such as YOLOv3, YOLOF, DeformDetr, SSD512, and SSDLit. Additionally, the training speed and detection accuracy are significantly improved in the surface defect detection of crazing (Cr), inclusion (In), patches (Pa), pitted surface (PS), polled-in scales (RS), scratches (Sc), and other surface defects, and the missed detection rate is greatly reduced.
Abstract: In the process of drilling, the speed at which a drill bit breaks through rock and deepens the drill hole is called the rate of penetration (ROP), which is an important index reflecting drilling efficiency. In recent years, machine learning methods have been applied to the ROP prediction. However, it is found in practice that the prediction accuracy of ROP based on existing machine learning methods is significantly reduced when applied to new oil fields, and the main reason is that the data available for learning and training in these new fields are usually scarce or even completely missing. Therefore, improving the prediction performance of ROP in new oil fields is an important issue to be solved. Considering this, a cross-oilfield ROP prediction method based on transfer learning is proposed, and a boosting transfer regression model with physical constraints is constructed to predict ROP of new oil fields. The experiments on real drilling datasets indicate that the proposed method is effective, and the prediction accuracy is significantly better than that of the current mainstream ROP prediction methods.
Abstract: A data association model and a mathematical motion simulation model are built for the real-time motion process of ground moving targets in ground support of airports. The simulation program is improved on the basis of the GIS geographic data, graphics device interface (GDI+), multithreading technology, and concurrent synchronization mechanisms. In the simulation process, the improved A* algorithm is used to determine the best driving path of moving targets at airports, and the moving states of the targets are monitored in real time using data visualization technology. Finally, the repetition method is employed to carry out multiple tests for statistical analysis of the simulations. The simulation model has been applied to many fields, such as the deduction of the ground support process of airports, airport emergency plan verification, and flight transportation decision-making. It is of great significance to improve the ability of airports to support flight transit.
Abstract: Although the attribution explanation method based on Shapley value can quantify the interpretation results more accurately, the excessive computational complexity seriously affects the practicality of this method. In this study, we introduce the k-dimensional (KD) tree to reorganize the predicted data of the model to be explained, insert virtual nodes into the KD tree so that it meets the application conditions of the TreeSHAP algorithm, and then propose the KDSHAP method. This method lifts the restriction that the TreeSHAP algorithm can only explain tree models and broadens the efficiency of the algorithm in calculating Shapley value to the explanation of all black-box models without compromising calculation accuracy. The reliability of the KDSHAP method and its applicability in interpreting high-dimensional input models are analyzed through experimental comparisons.
Abstract: Heart rate is an important physiological parameter for measuring human cardiovascular health and emotional stress. However, video-based non-contact heart rate detection techniques can degrade the detection accuracy in real scenarios due to facial movements and lighting changes. To solve the problem, this study proposes a new method of heart rate detection based on adaptive superpixel segmentation and multi-region integrated analysis depending on the high correlation between the selection of the region of interest (ROI) in a heart rate detection algorithm and its detection accuracy. Firstly, a face detection and tracking algorithm is used to crop the face image. Then the ROI is divided into non-overlapping sub-blocks by an adaptive superpixel segmentation algorithm. The original blood volume pulse matrix of each sub-block is constructed by chromaticity feature extraction. Finally, the pulse matrix is analyzed using multiple indicators, and the best region is selected for heart rate estimation. The experimental results show that the heart rate detection accuracy can be effectively improved by adaptive superpixel segmentation and optimal selection through multi-region analysis. The accuracy reaches 99.1% and 95.6% under stationary and motion disturbance conditions, respectively, and the accuracy is improved by up to 8.2% under illumination disturbance conditions compared with that of the traditional method. The proposed method enhances the robustness of heart rate detection in real scenarios.
Abstract: Designing and utilizing good image prior knowledge is an important way to enable image inpainting. A generative adversarial network (GAN) is an excellent generative model, and its generator can learn rich image semantic information from large datasets. Thus, it is a good choice to use a pre-trained GAN model as an image prior. Making use of multiple hidden variables, this study adds adaptive weights to the channels and feature maps at the same time in the middle layer of the pre-trained generator and fine-tunes generator parameters in the training process. In this way, the pre-trained GAN model can be used for better image inpainting. Finally, through the contrast experiment of image reconstruction and image inpainting and the combination of qualitative and quantitative analysis, the proposed method is proved effective to mine the prior knowledge of the pre-trained model, thus finishing the task of image inpainting with high quality.
Abstract: To address the problem of a small accepting neighborhood range during the node embedding of traditional graph convolutional networks, this study proposes a hyperspectral image classification network based on an improved GraphSAGE algorithm. Firstly, the original image is preprocessed by using the super-pixel segmentation algorithm to reduce the number of image nodes. This not only conserves the local topology information of the original image to the largest extent but also reduces algorithm complexity and thus shortened operation time. Secondly, the average sampling of the target node is carried out by the improved GraphSAGE algorithm, and the neighbor nodes are aggregated by the average aggregation function to reduce spatial complexity. Finally, the proposes approach is compared with other models on the public hyperspectral image datasets Pavia University and Kenndy Space Center. The experiment proves that the hyperspectral image classification network based on the improved GraphSAGE algorithm can achieve good classification results.
Abstract: In view of the problem that the traditional current protection method cannot be applied when distributed generation (DG) is connected to the distribution network, this paper takes the double-feeder distribution network line as the research object. Firstly, when three-phase short-circuit faults occur at different locations of the line, the influence of DG connected to a busbar of feeder ends or a non-end busbar on the short-circuit current flowing through each protection device in the line is analyzed. Then, a distribution network model is built by PSCAD software for simulation analysis. Since it is difficult to set the action value of short-circuit faults in the distribution network containing DG, a matrix algorithm based on intelligent electronic devices (IEDs) for fault information uploading is proposed, and the accuracy of the algorithm is verified by an example. The results reveal that when DG is connected to a busbar of feeder ends or a non-end busbar, the fault that occurs in the downstream of DG will cause maloperation of the protection device in the fault section, and the protection device in the upstream section may encounter operation failure, which is not conducive to fault positioning and removal. The proposed matrix algorithm is applicable to the distribution network with DG, and regardless of single or multiple faults, the fault area can be accurately located to ensure the safe and reliable operation of the distribution network.
Abstract: To tackle the problem that traditional container vector detection is limited to manual detection, this study designs a visual search system for container vectors based on machine vision. The system collects real-time video and captures the activity of vectors through a smart car under remote control. Then, it recognizes the vectors in the video returned by the car through deep learning and inter-frame detection. The system takes the YOLOv5 model as the training core and adopts a modular structure to realize the visual detection of container vectors. Machine vision helps improve detection efficiency and lays the foundation for the further use of robots to detect vectors.
Abstract: Air quality prediction is of great importance for people’s daily travel. As a new recurrent neural network (RNN) of deep learning, the long short-term memory (LSTM) network demonstrates good prediction ability for time sequence data. However, neural network models generally rely on experience for parameter selection during training and have a long training period, low prediction accuracy, and unreliable prediction results. Considering this, this study proposes a bidirectional LSTM model based on the whale optimization algorithm (WOA), namely, the WOA-BiLSTM model. Specifically, the BiLSTM network can enhance the memory capability of sequence data information by its forward and backward network structure, and WOA can assist the BiLSTM model in finding the optimal network parameters during the training process on the basis of the bubble-net hunting strategy of whales. The model is applied for air quality index (AQI) prediction in Shaanxi Province and compared with BiLSTM and LSTM models separately, and it is found that the proposed model registers the best prediction result with the MAE value of 6.543 3 and R2 value of 0.989 9. Therefore, the model is of good theoretical and practical significance for applications in air quality prediction.
Abstract: Currently, the target detection algorithm based on depth neural network is mostly used for pallet positioning and a rectangular box is generally utilized. The positioning accuracy of the pallet center point is not high enough, and the horizontal direction of the pallet cannot be estimated effectively. To solve this problem, this study proposes a pallet positioning method based on keypoint detection, which locates the pallet by detecting the four corners of the front outer outline. Firstly, due to the shortage of large-scale pallet data sets, the human posture estimation of CenterNet is introduced by transfer learning. Then the keypoint grouping method is improved, and the adaptive compensation is proposed for keypoint regression to improve the keypoint detection accuracy. According to the location of pallet keypoints, a method of pallet center point calculation and pallet horizontal direction estimation based on geometric constraints is proposed. Compared with the original CenterNet, the proposed method raises the positioning index APkp of pallet keypoint from 0.352 to 0.728, and the positioning accuracy ALP of pallet center point to 0.946. Meanwhile, it can effectively estimate the pallet horizontal direction and is of high practical value.
Abstract: Recently, the research on skeleton-based action recognition has attracted a lot of attention. As the graph convolutional networks can better model the internal dependencies of non-regular data, the spatio-temporal graph convolutional network (ST-GCN) has become the preferred network framework in this field. However, most of the current improvement methods based on the ST-GCN framework ignore the geometric features contained in the skeleton sequences. In this study, we exploit the geometric features of the skeleton joint as the feature enhancement of the ST-GCN framework, which has the advantage of visual invariance without additional parameters. Further, we integrate the geometric feature of the skeleton joint with earlier features to develop ST-GCN with geometric features. Finally, the experimental results show that our proposed framework achieves higher accuracy on both NTU-RGB+D dataset and NTU-RGB+D 120 dataset than other action recognition models such as ST-GCN, 2s-AGCN, and SGN.
Abstract: Federated learning protects user privacy by aggregating trained models of the client and thereby keeping the data local on the client. Due to the large numbers of devices participating in training, the data is non-independent and identically distributed (non-IID), and the communication bandwidth is limited. Therefore, reducing communication costs is an important research direction for federated learning. Gradient compression is an effective method of improving the communication efficiency of federated learning. However, most of the commonly used gradient compression methods are for independent and identically distributed data without considering the characteristics of federated learning. For the scene of non-IID data in federated learning, this study proposes a sparse ternary compression algorithm based on projection. The communication cost is reduced by gradient compression on the client and server, and the negative impact of non-IID client data is mitigated by gradient projection aggregation on the server. The experimental results show that the proposed algorithm not only improves communication efficiency but also outperforms the existing gradient compression algorithms in convergence speed and accuracy.
Abstract: Mobile edge computing (MEC) enables mobile devices (MDs) to offload tasks or applications to MEC servers for processing. As an MEC server consumes local resources when processing external tasks, it is important to build a multi-resource pricing mechanism that charges MDs to reward MEC servers. Existing pricing mechanisms rely on the static pricing of intermediaries. The highly dynamic nature of tasks makes it extremely difficult to effectively utilize edge-cloud computing resources. To address this problem, we propose a Stackelberg game-based framework in which MEC servers and an aggregation platform (AP) act as followers and the leader, respectively. We decompose the multi-resource allocation and pricing problem into a set of subproblems, with each subproblem only considering a single resource type. First, with the unit prices announced by MEC servers, the AP calculates the quantity of resources for each MD to purchase from each MEC server by solving a convex optimization problem. Then, each MEC server calculates its trading records and iteratively adjusts its pricing strategy with a multi-agent proximal policy optimization (MAPPO) algorithm. The simulation results show that MAPPO outperforms a number of state-of-the-art deep reinforcement learning algorithms in terms of payoff and welfare.
Abstract: Accurate named entity recognition is the basis of structured electronic medical records and plays an important role in the standardized writing of electronic medical records. However, current word segmentation tools cannot completely and correctly distinguish professional medical terms, making it difficult to achieve structured electronic medical records. As for problems in medical entity recognition, this study proposes an improved deep learning model base on BiLSTM-CRF in the field of named entity recognition. The model combines text and labels as input, which makes the model focus on more useful information in the multi-head attention mechanism. BiLSTM performs feature extraction on the input and obtains the probability of each text on all labels. CRF learns the constraints of the data set during the training and improves the accuracy of the results after decoding. The experiment uses 1 000 manually labeled electronic copies as the data set and the BIO for labeling. Compared with the traditional BiLSTM-CRF model, the proposed model raises the F1 value in the entity category by 3%–11%, verifying its effectiveness in named entity recognition of medical records.
Abstract: Financial institutions are currently grappling with the growth of non-performing assets (NPAs). The prediction accuracy of credit overdue directly determines the size of NPAs. For better prediction of repayment ability, data modeling methods are often introduced, which may cause over-fitting for new businesses with small data samples. This study performs case studies and enriches the small data samples by similarity with random forest, LightGBM, XGBoost, DNN, and TrAdaBoost transfer learning. It aims to provide an effective solution to insufficient samples during the model establishment for small sample businesses. The results show that the area under curve (AUC) of the five machine learning models is greater than 80 for small data samples after similar financial business data are integrated. The AUC of TrAdaBoost is at least 2 points higher than that of LightGBM, XGBoost, DNN, and random forest models on the prediction set. In addition, TrAdaBoost stands out with the highest precision (88%) and recall (73%).
Abstract: This study proposes an improved U-Net for precise segmentation of bone data to solve the problems of low contrast, indistinct features, and insufficient extraction of bone features by existing algorithms in bone computed tomography (CT) images. In the network coding stage, the densely connected dilated convolution module is used to enhance the extraction of bone features; in the network decoding stage, the attention-based fusion module is adopted to make full use of spatial information and semantic information and thereby avoid the loss of bone information. When the improved algorithm is applied to a CT dataset of human lower limb bones, the Dice coefficient is 89.44%, and the intersection over union (IoU) coefficient is 80.55%. Compared with those obtained with the U-Net model, the Dice coefficient is increased by 5.1%, and the IoU coefficient is improved by 7.63%. The experimental results show that the proposed optimization algorithm can be employed to achieve precise segmentation of CT images of lower limb bones. It also provides a reference for the preoperative planning for orthopedic diseases and subsequent treatment.
Abstract: As the basis of human motion recognition, two-dimensional human pose estimation has become a research hotspot with the popularity of deep learning and neural networks. Compared with traditional methods, deep learning can achieve deeper image features and express the data more accurately, thus becoming the mainstream of research. This study mainly introduces two-dimensional human pose estimation algorithms. Firstly, according to the number of people detected, the algorithms are divided into two categories for single-person and multi-person pose estimation. Secondly, the single-person pose estimation methods are divided into two groups based on coordinate regression and heat map detection. Multi-person poses can be estimated by top-down and bottom-up methods. Finally, the study introduces commonly used data sets and evaluation indexes of human pose estimation and compares the performance indexes of some multi-person pose estimation algorithms. It also expounds on the challenges and development trends of human pose estimation.
Abstract: Facial expression recognition (FER) has various applications in human-computer interaction scenarios. However, existing FER methods are not that effective for blurred and occluded expression. To cope with facial expression blur and occlusion, this study proposes a novel network based on local manifold attention (SPD-Attention), which uses manifold learning to obtain the second-order statistical information with a stronger descriptive ability for strengthening the learning of facial expression details and suppressing the influence of irrelevant features in the occlusion area on the network. At the same time, in view of the disappearance and explosion of gradient caused by logarithmic calculation, this paper proposes corresponding regular constraints to accelerate network convergence. The effect of the algorithm is tested on public expression recognition data sets, which is significantly improved compared with those of classic methods such as VGG. The accuracy is 57.10%, 99.01%, 69.51%, 87.90%, 86.63%, and 49.18% on AffectNet, CK+, FER2013, FER2013plus, RAF-DB, and SFEW, respectively. In addition, compared with state-of-the-art methods such as Covariance Pooling, the proposed method has an accuracy improved by 1.85% on a special blurred and occluded expression data set.
Abstract: As medical informatization is constantly improving, electronic medical records have been more and more widely used, of which the unstructured text contains massive measurable quantitative information including patient clinical conditions. Due to the complexity of entities and quantitative information, it is a challenge to accurately extract measurable quantitative information. In this study, we propose the RPA-GRU model combining the relative position feature and attention mechanism based on a bi-directional gated recurrent unit. It incorporates the relative position feature into the attention mechanism to identify entities and quantity information. Meanwhile, the GATM model is proposed according to the reconstructed dependency tree-based graph attention network to learn graph-level representation, thus achieving the association between entities and quantity information. The experiment is based on 1,359 electronic medical records from the burn injury department of a three-A hospital. The results show that the F1 values of RPA-GRU model and GATM model are 97.58% and 97.86% respectively in terms of identification and association of measurable quantitative information, up by 2.17% and 1.74% compared with the best-performing baseline model. In this way, the effectiveness of the proposed models is validated.
Abstract: Text similarity matching is the basis of many natural language processing tasks. This study proposes a text similarity matching method based on a Siamese network and character-word vector combination. The method adopts the idea of the Siamese network to model the overall text so that the text similarity can be determined. First, when text feature vectors are extracted, BERT and WoBERT models are used to extract character-level and word-level sentence vectors which are then combined to have richer text semantic information. If the dimension is too large during feature information fusion, the principal component analysis (PCA) algorithm is employed for the dimension reduction of high-dimensional vectors to remove the interference of redundant information and noise. Finally, the similarity matching result is obtained through the Softmax classifier. The experimental results on the LCQMC dataset show that the accuracy and F1 score of the model in this study reach 89.92% and 88.52%, respectively, which can better extract text semantic information and is more suitable for text similarity matching tasks.
Abstract: As imbalanced data are exposed to problems such as intra-class imbalance, noise, and small coverage of generated samples, an adaptive denoising hybrid sampling algorithm based on hierarchical density clustering (ADHSBHD) is proposed. Firstly, the clustering algorithm HDBSCAN is introduced to perform clustering on minority classes and majority classes separately; the intersection of global and local outliers is regarded as the noise set, and the original data set is processed after noise samples are eliminated. Secondly, according to the average distance between clusters of samples in minority classes, the adaptive sampling method with broader coverage is used to synthesize new samples. Finally, some points that contribute little to the classification of majority classes are deleted to balance the dataset. The ADHSBHD algorithm is evaluated on six real data sets, and the results can prove its effectiveness.
Abstract: Unknown malicious network traffic detection is one of the core problems to be solved in anomaly detection as the traffic data obtained from high-speed network data flow are often unbalanced and changeable. Although there have been many studies on feature processing and detection methods of unknown malicious network traffic detection, these methods have shortcomings in simultaneously solving data imbalance and variability as well as detection performance. Considering the difficulty in unknown malicious network traffic detection, this study proposes an unknown malicious traffic detection model based on integrated SVM and bagging. Firstly, in view of the imbalance of network traffic data, a traffic processing method based on Multi-SMOTE oversampling is put forward to improve the feature quality upon traffic processing. Secondly, considering the distribution diversity of network traffic data, an unknown traffic screening method based on semi-supervised spectral clustering is presented to screen unknown traffic from mixed traffic with a diverse distribution. Finally, with the idea of Bagging, an unknown malicious traffic detector based on integrated SVM is trained. The experimental results reveal that the proposed detection model is superior to the current similar methods in comprehensive evaluation (F1 value), and it also has good generalization ability on different data sets.
Abstract: Regarding the power system, front-end technologies are diverse at present, whereas the manual coding is inefficient and cannot meet the rapid growth of demand. In response, this study designs a low-code visualization page editing engine for the power system with visualization and virtual DOM technologies in light of the component-based design idea. The virtual DOM technology is applied to the high-performance rendering of most scenes during page building. A unified data model is designed to integrate heterogeneous data and share data among components. The idea of a multi-type template for page editing is proposed to meet the requirement of the business system for diversified integration. The practice indicates that the system can provide agile, efficient, and low-code development under low barriers, which significantly improves the development efficiency of front-end pages.
Abstract: Fuzzing is outstanding in detecting vulnerabilities in real-world programs. In recent years, researchers have paid considerable attention to fuzzing improving techniques, and large numbers of optimized fuzzers were proposed. These fuzzers are usually combinations of more than one improving technique for better performance. However, systematic evaluation of individual fuzzing improving techniques is still to be conducted. In this study, we established an evaluation system for such techniques according to four metrics and used it to evaluate individual fuzzing improving algorithms integrated into recently proposed advanced fuzzers. Multiple groups of experiments were conducted to evaluate the effectiveness of each individual technique in each category of improving techniques, and the experimental data were comprehensively analyzed with the actual algorithm design and code implementation. We hope the evaluation of individual fuzzing improving techniques could help researchers develop more effective fuzzers in the future.
Abstract: For the privacy leakage during the data collection of fog-assisted smart grids, this study proposes a novel privacy-preserving data aggregation scheme with fault tolerance. Firstly, the BGN homomorphic encryption algorithm and the Shamir secret sharing scheme are combined to protect data privacy. At the same time, an efficient signature authentication method is constructed based on the elliptic curve discrete logarithm problem to ensure data integrity. In particular, the scheme has two fault-tolerant measures. When some smart meter data cannot be sent normally or some cloud servers fail to work because of attacks, the scheme can still perform aggregate statistics. The security analysis proves that the scheme meets the security requirements of the smart grid. The performance experiments show that the proposed scheme has better computational and communication performance than the existing schemes.
Abstract: The security of electric energy plays an important role in national security. With the development of power 5G communication, a large number of power terminals have positioning demand. The traditional global positioning system (GPS) is vulnerable to spoofing. How to improve the security of GPS effectively has become an urgent problem. This study proposes a GPS spoofing detection algorithm with base station assistance in power 5G terminals. It uses the base station positioning with high security to verify the GPS positioning that may be spoofed and introduces the consistency factor (CF) to measure the consistency between GPS positioning and base station positioning. If CF is greater than a threshold, the GPS positioning is classified as spoofed. Otherwise, it is judged as normal. The experimental results show that the accuracy of the algorithm is 99.98%, higher than that of traditional classification algorithms based on machine learning. In addition, our scheme is also faster than those algorithms.
You are the firstVisitors
Copyright: Institute of Software, Chinese Academy of Sciences Beijing ICP No. 05046678-3
Address：4# South Fourth Street, Zhongguancun,Haidian, Beijing,Postal Code：100190
Phone：010-62661041 Fax： Email：csa (a) iscas.ac.cn
Technical Support：Beijing Qinyun Technology Development Co., Ltd.